Review

International Journal of Distributed

Sensor Networks
2021, Vol. 17(10)
Review and classification of trajectory Ó The Author(s) 2021
DOI: 10.1177/15501477211050729
summarisation algorithms: From journals.sagepub.com/home/dsn

compression to segmentation

Daniel Amigo , David Sánchez Pedroche , Jesús Garcı́a and

José Manuel Molina

Abstract
With the continuous development and cost reduction of positioning and tracking technologies, a large amount of trajec-
tories are being exploited in multiple domains for knowledge extraction. A trajectory is formed by a large number of
measurements, where many of them are unnecessary to describe the actual trajectory of the vehicle, or even harmful
due to sensor noise. This not only consumes large amounts of memory, but also makes the extracting knowledge pro-
cess more difficult. Trajectory summarisation techniques can solve this problem, generating a smaller and more manage-
able representation and even semantic segments. In this comprehensive review, we explain and classify techniques for
the summarisation of trajectories according to their search strategy and point evaluation criteria, describing connections
with the line simplification problem. We also explain several special concepts in trajectory summarisation problem.
Finally, we outline the recent trends and best practices to continue the research in next summarisation algorithms.

Keywords
Trajectory summarisation, trajectory segmentation, trajectory compression, data compression, Douglas–Peucker, spatial
data analysis, trajectory partitioning

Date received: 19 April 2021; accepted: 9 September 2021

Handling Editor: Lyudmila Mihaylova

Introduction endangered animals, and to everyday and everyone

functions such as the use of the Global Positioning
Geolocation is a technique that makes possible to give System (GPS) navigation system in their cars (164 mil-
a position to an object by identifying its geographic lion people in the United States use it in their mobile
position on the Earth at a moment in time. It can be phones).
achieved by external sensors that allow tracking (radar, This increase in existing information related to geo-
Light Detection and Ranging (LiDAR) and video) or location allows it to be exploited using data analysis
using internal sensors (global navigation satellite sys- approaches, like machine learning and big data, making
tem (GNSS)) that achieve their own geolocation.
It is a technique that has existed since the 1950s in
the military and space fields. Nowadays, it is accessible Applied Artificial Intelligence Group (GIAA), University Carlos III of
to everyone in tiny devices with high precision at low Madrid, Madrid, Spain
consumption and manufacturing costs. This has pro-
gressively made the applications of the technology Corresponding author:
Daniel Amigo, Applied Artificial Intelligence Group (GIAA), University
spread to all sectors: from military tasks such as pre- Carlos III of Madrid, Avenida de Gregorio Peces-Barba Martı́nez, 22,
cisely locating the position of a fighter jet, to transport Colmenarejo, Madrid 28270, Spain.
uses like monitoring cargo shipments or surveillance of Email: [email protected]

Creative Commons CC BY: This article is distributed under the terms of the Creative Commons Attribution 4.0 License
(https://creativecommons.org/licenses/by/4.0/) which permits any use, reproduction and distribution of the work
without further permission provided the original work is attributed as specified on the SAGE and Open Access pages
(https://us.sagepub.com/en-us/nam/open-access-at-sage).
2 International Journal of Distributed Sensor Networks

it possible to obtain new knowledge. It can be applica- from this summary moving from these compression
ble at different levels to improve and refine intelligent techniques towards trajectory segmentation techniques
systems. that summarise them using segments that are represen-
By grouping geolocation measurements of the same tative of the different parts of the trajectories and pro-
object ordered in time, it is possible to generate trajec- vide a semantic description. This trend does not negate
tories that represent the movement of the geolocated older techniques that only sought to compress as there
object. According to Zheng,1 there are four types of are many approaches that use compression techniques
trajectories depending on the object that perform the to obtain representative segments.
trajectory (people, vehicles, animals or natural Advancing from the more simplistic approach of the
phenomena). line simplification problem there is the inclusion of time
It is estimated that by 2022 there will be 29 trillion within the data, which allow the work in time series
connected devices in the world, with more than 62% of and trajectory information. In the 2000s, this transition
them being related to the Internet of Things (IoT).2 started with Keogh et al.4 segmenting time series,
Among today’s devices, GPS typically has a refresh Meratnia and De By5 introducing the time dimension
rate of 10 Hz, which means that trajectories of long in the compression process, and Anagnostopoulos
duration can become very heavy. In fact, one experi- et al.6 starting the trajectories segmentation.
ment3 proved that storing the GPS records of 400 cars There are reviews and surveys that cover this prob-
monitored throughout the day costs approximately lem of summarisation in the literature,7–14 although
100 MB per day. most reviews address this problem in a tangential way
This implies that the available trajectory information as they focus on more generic problems. The ones that
is enormous, meaning a time and processing capacity explore this problem are brief and leave aside the com-
cost that may be too high. In addition, the use of partic- pression or segmentation branches. Moreover, they
ularly long trajectories can result in failures of the data explain the problem from the time series or trajectories
mining techniques due to the inability to analyse the point of view exclusively, without addressing the con-
details of the trajectory. One recommendable approach nection and distinction between the two approaches.
in data mining of trajectories is the decision into smaller Since there are so many different approaches to the
parts (segments) simplifying the search for patterns of trajectory summarisation problem and the absence of a
interest.1 review that covers them all from a point of view that
Hence, it appears the need to summarise trajectories summarises the whole spectrum, in this article a review
in a way that makes the information stored in the tra- and classification of the literature is attempted. This
jectories more processable and useful. This new term article considers the whole spectrum of trajectory sum-
introduced in this work covers a whole spectrum of marisation, focusing on compression and segmentation
other closely related terms, such as trajectory compres- techniques.
sion or trajectory segmentation. When summarising This study of the literature has collected 162 sum-
trajectories, the simplest approach is data compression, marisation algorithms. All of them have been analysed
specifically trajectory compression, which seeks to and classified within parameters extracted after the
reduce the amount of data stored to obtain a trajectory analysis of each paper and algorithm design that cover
with less weight when it is processed, stored or sent, both segmentation and compression algorithms. In
reducing costs in each aspect and speeding up any tra- addition to algorithms classification, the tests proposed
jectory processing algorithm. for each algorithm are studied to check whether their
Trajectory compression algorithms stem from line performances have been proven to be predominant in
simplification algorithms. With the ‘birth’ of comput- the literature. This complete study of each algorithm
ing, there were many uses of vectorial figures: represen- can be found and download at the website:15 https://
tation of maps, drawings for printing and so on. The danielamigo.github.io/trajectorySummarisationReview/.
computational constraints were much greater than they It allows to compare these algorithms through the para-
are today. As a result, the various tasks could not deal meters, thus better understanding their similarities and
with high-resolution data, making it necessary to sim- differences.
plify the lines and polygons used. Researchers as It has been observed that, although these two
Bellman, Douglas–Peucker, Jenks, McMaster, and approaches to the problem of summarising trajectories
many others addressed that problem. exist (compression and segmentation), the algorithms
Today, such computational limitations are not that used for both approaches are similar and have many
relevant, and the problem has changed by having an characteristics in common, being two of the most
additional dimension with the timestamp of the trajec- important characteristics:
tory. Therefore, the current trend in the literature is no
longer to summarise trajectories to reduce the storage  The search strategy consists of the methodology
resources. The objective is to discover new knowledge used to study the whole set of all raw trajectory
Amigo et al. 3

points. Depending on the strategy, a higher or  Algorithms with probabilistic models of the tra-
lower quality representation can be obtained, jectory movement are promising solutions. For
but it will affect the computation time needed to example, self-adaptive online trajectory sampling
obtain the summarisation. (SAOTS) or interacting multiple model (IMM)
 The evaluation criteria which are the method provide good results. Note also that they are
used to evaluate whether each subset of the capable of producing semantic content.
points studied by the strategy should belong to Alternatively, without modelling their dynamics,
the raw trajectory. This preservation criterion window strategy–based algorithms such as
gives priority to one type of result in terms of SQUISH-E or opening window-time ratio
the summarisation to be obtained, so it is impor- (OPW-TR) perform well, achieving a good bal-
tant to choose it according to the problem to be ance of computational cost with easy tuning
solved. parameters.
 If it is not required a real-time operation, batch
During the literature study, certain special algo- solutions are preferable to online solutions.
rithms were found that approximate the problem with Among this type, the ones that perform a graph-
unique characteristics. For example, there are lossless based strategy stand out. Algorithms such as
compression algorithms that are focused on not losing directed acyclic graph based online trajectory
information when summarising the trajectory, or algo- simplification (DOTS) or multiresolution poly-
rithms to summarise trajectories considering the road gonal approximation (MRPA) obtain subopti-
networks on which they move. We also find summarisa- mal solutions with reasonable computation
tion algorithms aiming to generate knowledge directly. times.
Known as semantic summarisation, they generate seg-
ments with a specific behaviour. This behaviour can be The main contributions of this review can be sum-
related to the movement dynamics, for example, high-, marised by the following aspects:
or low-speed segmentation, or related to the context,
for example, stopping near a specific location.  An introduction and motivation of the trajectory
In addition, our literature review pointed out other summarisation problem and its links with trajec-
common characteristics, showing a trend change over
tory segmentation and compression techniques.
the years in the algorithms, which should prevail in  An accessible global classification of all types of
future works. For instance, the shift from the data used
trajectory summarisation, focusing in two
for the summarisation, adding other dimensions like the
aspects: the search strategy and the evaluation
temporal data, the need to obtain the summarisation
method for selection of key points.
quickly or even in real time, or the search of the best  A compilation of notable approaches found in
parameters of the algorithms to obtain good results.
the literature for specific sets of algorithms.
This work does not intend to conclude which algo-  A compilation of common features to all algo-
rithm is the best for each use, as it is an impossible
rithms found in the literature, introducing
task. It is necessary a specific analysis depending on the
important trends to preserve in future works.
intended use and data characteristics to decide the best
one according to the needs of each problem: online or
batch compression, limited computational power, The remainder of this article continues as follows.
mobility constraints such as roads, prioritisation of Section ‘Basic concepts’ introduces some basic concepts
other variables such as orientation or semantic content of trajectory summarisation to fully grasp the rest of
and so on. Still, one way to identify how good is a par- the work. Section ‘Trajectory summarisation algo-
ticular algorithm is to check its paper’s comparisons rithms’ provides the two main categories proposed to
with other algorithms (column ‘Comparison to other classify all the algorithms reviewed. Section ‘Special
algorithms’ on the website15). In order to facilitate the approaches’ describes several special approaches for
navigation through the many algorithms, a brief and trajectory summarisation, while section ‘Other charac-
general summary of the overall findings of this work is teristics’ discusses other secondary classifications to
provided as follows: highlight trends to be followed in future works. Finally,
section ‘Conclusion’ concludes the work.
 Overall, this study concludes that the traditional
line simplification algorithms, such as the
Basic concepts
well-known Douglas–Peucker algorithm, are
outdated for trajectory summarisation, as there In this section, some preliminary concepts are intro-
are plenty alternatives that provide improved duced and formally defined to understand the follow-
results across all metrics. ing sections of this article. Table 1 summarises all the
4 International Journal of Distributed Sensor Networks

Figure 1. Trajectory example.

Definition 4 (raw trajectory). Original trajectory before

Table 1. Notation summary.
any processing is represented as T = fP1 , P2 , . . . , Pn g.
Notation Definition
Definition 5 (summarised trajectory). A summarised trajec-
T Raw trajectory tory is a trajectory formed by a subsequence of the tra-
T0 Summarised trajectory jectory points (selected trajectory points in Figure 1) of
Pt Trajectory point, represented by the spatial
a raw trajectory. It is represented as
information and the associated time
(xt , yt ) Spatial information of a trajectory point on local T 0 = fP1 , P2 , . . . , Pm g, where T 0  T . To obtain the tra-
coordinates jectory point subsequence, it is necessary to use a sum-
ƒ!
Ps Pe Segment that starts in Ps trajectory point and marisation algorithm.
ends in Pe trajectory point
Pt0 Trajectory point projection Definition 6 (segment). A segment is a subtrajectory
formed by two consecutive points of a summarised tra-
jectory. For example, in Figure 1 trajectory, P1 and P4
ƒƒ!
notations presented in the section. The concepts are form the segment represented as P1 P4 . It summarises
explained supported by the illustration of Figure 1. the associated points of the raw trajectory, which are
the ones that are located between P1 and P4 .

Definition 1 (time series). A list of ordered tuples, being Definition 7 (trajectory point projection in
one part of the tuple, the time reference corresponding segment). Represented as P0n , is the representation of
to the measure magnitude. The other part of the tuple non-selected trajectory point (Pt ) on its associated seg-
is the measurement itself, which varies according to the ment. In example, P02 is the projected point in segment
problem. ƒƒ!
P1 P4 of trajectory point P2 .

Definition 2 (trajectory). Time series that stores target Definition 8 (compression ratio). A ratio that measures
localisation data over time. The second part of the tuple how much a summarised trajectory is reduced with
is the measurement of the target position at each time respect to the raw trajectory. It is measured by dividing
instant. the number of removed points of the raw trajectory to
form the summarised trajectory with respect to the total
Definition 3 (trajectory point). A trajectory point is a tuple points of the raw trajectory. In Figure 1 trajectory, it is
that stores the measurement of the target at a certain 1  (4=11) = 64%.
time. Therefore, a trajectory point is formed by two
components: the timestamp when the measurement was Definition 9 (semantic trajectory). Summarised trajectory
taken and the spatial location of the target in that time. in which each of the segments has a semantic meaning
The spatial information can be represented in local specific to the problem, for instance, uniform, turn,
(xt , yt ) or global (Latt , Lont ) coordinates may have a stop and so on.
third dimension (zt or Heit ) if the points form a three-
dimensional (3D) trajectory. It is represented as Definition 10 (summarisation algorithm). Algorithm used to
pt = (timet , spatialt ). obtain a summarised trajectory by the processing of a
Amigo et al. 5

raw trajectory. It needs a search strategy to process the

trajectory points sequence and an evaluation criterion
that decides if each point should be in the summarised
trajectory subsequence.

Definition 11 (evaluation criteria). The criteria that any

summarisation algorithm has. Is used to decide if a tra-
jectory point should be included in the summarised tra-
jectory subsequence or not.

Definition 12 (search strategy). Methodology that differs

between the different algorithms and is used to pass
over all the raw trajectory points making the process of
the entire sequence.

Trajectory summarisation algorithms

As already indicated, the algorithms that summarise tra-
jectories have the objective of calculating the most rele-
vant points of a raw trajectory to obtain a summarised
trajectory. In the whole set of algorithms, two key ele-
ments have been found by means of which it is possible
to classify the different algorithms, the search strategy
and the evaluation criteria to select the key points.
Therefore, to summarise the different algorithms
analysed, this section is broken down into two sections:

 The first one focuses on the relevant point selec- Figure 2. Summarisation algorithms classification.
tion criteria, which summarises the different
approaches found when deciding whether to
keep or not to keep each point in a subsequence
the raw trajectory should belong to the summarised tra-
of the raw trajectory within the summarised
jectory. This process is commonly referred to as heuris-
trajectory.
 tics. On the simplest criteria it might seem appropriate,
The second one consists of the processing strat-
although it is not for more complex approaches that are
egy, and summarises the different approaches
being developed.
found when processing the set of points of the
This selection of points is usually done by giving a
trajectory to evaluate the subsequence to be sim-
specific score to each trajectory point. This score is
plified based on the selection criteria.
based on a specific methodology to quantify by means
of a concrete analysis how good a point is compared to
Note that these two concepts are not separated but
another. The strategy will use this score to decide at
act in tandem to form the algorithm that finds the sum-
marised trajectory. each moment which point should be included in the
Figure 2 resumes the different possible classifications summarised trajectory and which point should be
that have been found within these two main categories, discarded.
a trajectory summarisation algorithm may work by Throughout the literature review, it has been
combining a strategy with a point selection criterion. observed that this category groups the algorithms into
In each of the following sections, only the most rele- the following subcategories according to the methodol-
vant algorithms will be mentioned. At the end of the ogy used for the selection of these representative
section, Table 2 indicates where in these categories each points:
of the algorithms studied belongs.
 Trivial: the most basic algorithm approaches.
They do not use any score, only rely on a very
Trajectory point evaluation criteria basic rule to make the inclusion decision.
As mentioned previously, all trajectory summarisation  Distance: these algorithms use the distance
algorithms need a method to decide whether a point in between relevant points in the summarisation
6 International Journal of Distributed Sensor Networks

Table 2. Trajectory summarisation table.

Strategy Preserve Techniques

criterion

Sequential Distance AMS,16 TD-TR reduce,17 WKMeans,18 Pyramid,19 ADP,20 CB-SMoT,21 STC,22
GRASP-UTS,23 RGRASP-SemTS,24 BTC,25 OLDCAT,26 SGTCR-CS,27 GSC and GSTC28
Angle TD-TR reduce,17 Persistence,29 OLDCAT26
Velocity TD-TR reduce,17 CB-SMoT,21 AACAT,30 SimpleTrack,31 SAS,32 SAOTS,33 OLDCAT26
Transform Coresets,34 AACAT,30 SimpleTrack,31 SGTCR-CS27
Probabilistic IMM,35–37 APSOS,38 SAS,32 SAOTS,33 SGTCR-CS27
Graph Distance Bellman,39,40 DOTS,41 DOTS-CASCADE,41 Iri-Imai,42,43 MRPA,44 Daescu,45,46 OGPC
and OSPC,47 MMTC-offline,48 MMTC-online,48 SPPA,49 GRTSOpt,50 Latecki,51 Trajic,52
Representativeness,53 KAA and StreamKAA,54 OLTS and OPTTS,55 DOTS*,56 OSC and
OSTC,28 CLEAN57
Angle VTracer,58 DPTS + ,59 Latecki,51 SP60
Velocity DOTS*56
Opening Window Distance PoI and PoIE,61 GRPPA,62 TSHL,63 AMS,16 CFF,64 BOPW and NOPW,4 OHTA,
OnlineOHTA and SATA,65 CDR, CDRm, GRTSOpt and GRTSSec,50 TraClus,66 OPERB
and A-OPERB,67 BQS,68 ABQS, FBQS and PBQS,69 LO-OPW-TR,70 OPW-TR,3 SMoT,71
Pan,72 Patroumpas,73,74 STTrace,75 Resheff,76 Reumann-Witkam,77 EPP,78 SplitTrajs,79
BTC and HTC,25 TPMF,80 DR,81 SetraStream,82 ROCE,83 SPD84
Angle GRPPA,62 TSHL,63 CFF,64 Angular,85 Interval,86 OHTA, OnlineOHTA and SATA,65
TraClus,66 OPERB and A-OPERB,67 BQS,68 ABQS, FBQS and PBQS,69 Intersect,60
Error-Search, Min-Error and Span-Search,87 Pan,72 Patroumpas,73,74 Thresholds,75 EPP,78
SplitTrajs,79 BTC,25 TPMF,80 Zhao-Saalfeld88
Velocity SUTC,89 OPW-SP,3 Pan,72 Patroumpas,73,74 Thresholds,75 SplitTrajs,79 TPMF,80
CoTracks90
Probabilistic TSHL,63 OPERB,67 A-OPERB67
Sliding Window Distance SWS,91 OWS,92 WSII,93 ISW,94 Opheim-improved,95 RSLC and TSLC,96 FFUS,97 FSW,98
BQS,68 ABQS, FBQS and PBQS,69 FastSTray,99 TD-TR,3 SQUISH,100 SQUISH-E(l) and
SQUISH-E(m),101 Opheim,102 STMaker,103,104 ISW-SPM,105 TSA1 and TSA2,106 DPSW107
Angle SWS,91 OWS,92 WSII,93 ISW,94 DPBGD,86 FFUS,97 GS,108 BQS,68 ABQS, FBQS and
PBQS,69 FastSTray,99 Pikaz,109 STMaker,103,104 DPSW107
Area VW-TS,110 VW111
Velocity GS,108 TD-SP,3 DPSW,107 HESAVE112
Transform FastSTray99
Split Distance GRPPA,62 Similarity,113 DP,114 SWS,91 OWS,92 WSII,93 TCMM,115 TD-TR reduce,17 DP-
hull,116 SWAB,4 FFDP,97 GRTSSec,50 Pyramid,19 ATS,117 ADP,20 INCM,118 ESTC-EDP,119
SPM,120 STMaker,103,104 ISW-SPM,105 SELF,121 TPMF,80 DPSW,107 SNDSC,122 SPM3D,123
VO124
Angle GRPPA,62 SWS,91 OWS,92 WSII,93 TCMM,115 TD-TR reduce,17 DPDP,86 FFDP,97 ATS,117
STMaker,103,104 SELF,121 TPMF,80 DPSW,107 SNDSC122
Area IC-MBR125
Velocity 2stage-pls,126 TCMM,115 TD-TR reduce,17 ESTC-EDP,119 SELF,121 TPMF,80 DPSW,107
SNDSC122
Merge Distance TS,127,128 SWAB,4 DMin, S-DMin and SE-DMin,129 SQUISH,100 SQUISH-E(l) and
SQUISH-E(m),101 STTrace,75 GSC and GSTC28
Angle TS,127,128 Persistence,29 DMin, S-DMin and SE-DMin,129 GS108
Area Anagnostopoulos,6 EXTA,62 IC-MBR,125 Pikaz,109 VW-TS,110 VW111
Velocity GS108
Probabilistic TS127,128

process or the trajectory to make the preserve  Area: these algorithms calculate areas by mer-
decision. ging several points in the summarisation process
 Velocity: these algorithms use the velocity in to make the decision.
points to make the decision.  Transform: these algorithms are based on the
 Angle: these algorithms use the angle difference definition of a series of points that mathemati-
between several trajectory points to make the cally generate a function that approximates the
decision. trajectory.
Amigo et al. 7

 Probability: these algorithms use probabilities points is not as similar to the raw trajectory as desired.
calculated by the algorithm itself to make the Also, it can be used to choose which point in the trajec-
decision. tory should be preserved. By carrying out this process
 Based on multiple criteria: these algorithms com- through all the trajectory points, the strategy will find
bine several of the above criteria to make the the summarised trajectory.
decision. The first and logical version of this distance uses the
shortest path from the trajectory point to the segment.
Trivial. Of all the ways of point selection, this is the sim- This distance is the Euclidean distance, known in the lit-
plest possible. Unlike the other methods, this method erature as Perpendicular Euclidean Distance (PED). It
does not perform any analysis of the trajectory charac- was first introduced with the best-known algorithm in
teristics to select the trajectory point to preserve. the summarisation literature, Ramer–Douglas–Peucker
Instead, it is based solely on a simple selection criterion (DP). Initially proposed by Ramer133 in 1972 and refined
applied to a list of points. by DP,114 the DP algorithm makes splits in the trajec-
The first solution in this classification is known as tory by the trajectory point with the highest PED.
the nth point routine or uniform sampling. Much of the As this metric is measured at trajectory point level,
literature gives this application to the work of it can be used in multiple ways. DP uses the maximum,
Tobler.130,131 In this, points are selected with a constant but other researchers use it in a grouped form over
sampling of N measurements, discarding for summari- time, with the metrics Integral Square Error (ISE)134
sation the N 2 1 measurements in between two selected and Local Integral Square Error (LISE).135 ISE quad-
measurements. In this way, a specific compression ratio ratically groups all the PED distances of the trajectory.
is ensured, and segments of a fixed size are obtained. It has a high computational cost but ensures an optimal
The other solution found in the literature, instead of solution. However, LISE only accumulates the errors
relying on a uniform criterion, evaluates each point ran- of the current segment, ignoring the rest of the seg-
domly. On each trajectory point, it applies a random ments. Therefore, its solution will be suboptimal,136
function to decide whether to keep that trajectory point although it has a better computational cost.
or not. Vitter132 is the reference that encompasses these This whole process was designed for line simplifica-
approaches. He made a proposal and comparison of tion solutions. From the 2000s onwards, when trajec-
line compression using reservoir sampling. tories became popular, researchers realised that current
This type of algorithm has the advantage of having algorithms, designed for geometric shapes, were not
a very low execution cost, making it a very simple and valid for trajectories.3 Trajectories are not merely a spa-
fast way to generate a series of segments. Conversely, if tial shape but had an extra dimension with the time at
these segments have a high level of compression, they which each trajectory point is measured.
will lose the most complex and sharp parts of the tra- Meratnia and De By3 introduced a way of introdu-
jectory, which is a big drawback for future analyses. cing the time dimension into the preserve criterion,
using its Time Ratio (TR) metric. Instead of calculating
the distance of the raw point perpendicular to the seg-
Distance. As mentioned previously, trajectory compres- ment, it performs a projection of the actual point onto
sion algorithms naturally emerge from the polygonal the segment. This projection is calculated by adjusting
approximation and line simplification algorithms. The the time travelled on the raw trajectory and the dis-
data used by these algorithms consisted only of ordered tance, compared to the distance of the summarised seg-
geometric points which, connected by lines, form fig- ment. This makes the projected trajectory point
ures or polygons. maintain the time proportions even on the segment.
This approach is therefore the most common Later, Potamias et al.75 made a metric with the same
throughout the literature, because of the clear impor- objective but more efficient, called Synchronous
tance of a trajectory shape over the plane. The distance Euclidean Distance (SED). The latter is widely accepted
is used to compare two points with each other in the by many researchers. The difference between PED and
same coordinate system. In this problem, distance can SED distance can be seen in Figure 3.
be applied to different relevant points, each one being a As with PED, there are the cumulative metrics ISE
different approximation. and LISE, algorithms such as MRPA44 and DOTS41
adapt them to SED with integral square synchronous
Trajectory point and segment distance. The most com- euclidean distance (ISSD) and Local Integral Square
mon use of distance is the comparison between the raw Synchronized Euclidean Distance (LISSED), respectively.
trajectory and the summarised one. For each point on
the raw trajectory, the distance to the summarising seg- Consecutive trajectory points distance. Another way to
ment can be measured. This distance can be used to use distance is to measure the separation between con-
detect if the summarised segment after removing some secutive trajectory points, as it is represented in grey in
8 International Journal of Distributed Sensor Networks

This evaluation method allows to focus the summar-

isation process on the preservation of the most delicate
components of a trajectory, such as the sharpest angles.
There are approaches, commonly called the Direction
Preserve Trajectory Simplification (DPTS), first intro-
duced by Long et al.,60 which aim to provide the best
heuristics to store this useful information about the
vehicle dynamics. As happens in the distance-based
metrics, this metric is also applicable to line simplifica-
tion problems, due to the lack of a time component.
Figure 3. Trajectory example with PED, SED and consecutive Due to low precision of trajectory data, noise can
distance. generate angles that are sharper than they really are.
This noise is minimised using the already seen distance
Figure 3. This value is used by several researchers to metrics but can affect this direction preserving algo-
check whether one measurement and the next one are rithms, preserving such noisy measurements, and dis-
separated within a suitable range. If not, it will be nec- carding the real motion ones. Some specific direction
essary to create a new segment. preservation techniques that take this noise into
Resheff76 does a version of maximum distance account when dealing with the angles.
between segments radially, integrating also the density Long et al.60,87 proposed several optimal and subop-
of nearby points. Sheng et al.79 make the same type of timal simplification algorithms. Latecki and
radial distance in a maritime environment. Lakämper51 calculate the difference of angles between
Opheim102,137 does something similar, but generates a the previous and the current direction vector. In this
rectangular area with radial corners, in which, if the case the data are points of a line simplification, to pre-
following points are inside, they will be compressed. serve the shape and not to blur the edges.
A middle term between the two distance approaches Wang et al.139 use the angle formed by three consec-
is created by Dead Reckoning, proposed first by utive trajectory points. The angle of the intermediate
Trajcevski et al.81 Instead of measuring the distance position with respect to the other two, called by them
from the trajectory point to the segment, they establish open angle, when it is an angle far from 180 degrees,
a predictive zone where the next trajectory point should represents a sharp turn that must be stored in the sum-
enter, following the trend of the previous ones. marisation, and calculates the difference of angles
Reumann and Witkam77 had previously defined a simi- between the previous and the current direction vector.
lar concept, where two parallel lines delimit the possible In this case the data are points of a line simplification,
position of the next one to maintain the current to preserve the shape and not to blur the edges.
segment. Ke et al.85 propose a grouping of the difference of
Another way to use this distance between consecu- the angular values of the vectors, so a change of seg-
tive trajectory points is to measure the accumulation ment is applied if several trajectory points show a
between several of them, taking the length of a given course change by the comparison with a threshold. An
trajectory. This value can be used to compare, as Cui example of this algorithm can be seen in Figure 4.
et al.63 or Sheng et al.79 do with a maximum segment Katsikouli et al.29 perform a different approach, detect-
length limit. ing local maximum and minimum angles over time in a
Finally, there are several approaches that use the way that preserves them.
previous distance concepts but measuring other types
of points, which are auxiliary to the trajectory or to the
summarisation process. They use to be relevant geogra- Area. Other method brought directly from line simplifi-
phical places for the analysis to be carried out, such as cation is the calculation of the area (or a volume if has
road intersections,25 regions of interest near the trajec- three dimensions) formed by a group of points of the
tory138 or even other trajectories.53 With the knowledge summarisation process.
of this distance, semantic content can be generated to Visvalingam and Whyatt111 consider this metric to
be exploited in the future. be more reliable for this type of problem, seeing it as a
grouping of distance and angle. Only those that are fea-
sible according to the angle they form (feasible) enter
Angle. The raw geometric representation of trajectory the network. However, it has the disadvantage of being
points allows the generation of more types of metrics a somewhat more complex and costly calculation to
to be used during the summarisation process. The angle generate.
formed by the relevant trajectory points when summar- It should be noted that the distance metrics that cal-
ising can be compared with others along the trajectory. culate a region in which the next trajectory point must
Amigo et al. 9

Figure 5. Visvalingam–Whyatt area illustration.

Figure 4. Example of accumulated angle criterion.

acceleration or angular velocity and segment accord-

be located are not area metrics. Those regions are visual ingly. In addition, it allows segments to be generated
representations to observe if the distance is fulfilled or that, for example, exceed the maximum speed of the
not, while these calculate a volumetric value to compare area where they are moving, generating useful context
with a threshold. in other applications.
In a similar way to the angle differences, Pikaz and When Meratnia and De By3 introduced the distance
Dinstein109 use three consecutive trajectory points to TR metric, also created the algorithm Opening Window
calculate the area of the triangle formed between them Spatiotemporal. This algorithm, in addition to using
as a decision criterion. The biggest area central point is TR, checked a maximum speed differential to perform
eliminated until a threshold is reached. This process is segmenting, since the vehicle when accelerating or
illustrated in Figure 5. A similar variant performs decelerating is in another type of motion. Potamias
Visvalingam and colleagues.111,140 Later, a modification et al.75 introduced the algorithm thresholds. It makes a
of the Visvalingam–Whyatt algorithm that includes prediction of the zone in which the next trajectory point
time in the calculation of the area was proposed.110 should appear, but specifically takes the speed into
Another common area approximation is to perform
account in the calculation of the valid zone. This
Minimum Bounding Rectangles (MBR). These are
ensures that in addition to following the same direction,
areas formed by rectangles on a plane, without rota-
a constant velocity is maintained.
tion. The goal of these approaches is to encapsulate all
This is also done using the distance, and the
trajectory points in the fewest number of MBRs, mini-
trajectory-point velocity can be compared with the
mising the total area. Liu et al.125 apply it to the plane
equivalent point projected on the summarised trajec-
while Anagnostopoulos et al.6 do it in a volumetric
tory segment. De Vries and Van Someren126 use this
way, introducing time as another dimension.
approach to detect movements and stops, making seg-
The MBRs are a simplified form to calculate the
ments to represent this movement.
area, since only the base and height of the rectangle are
Another example of the use of speed in summarisa-
needed. Others62,125 have proposed the metric that rea-
tion is that proposed by Lin et al.117 It uses the Gini
lises the area between the raw trajectory points and
index on the velocity values to detect the points at
their projection on the segment of the summary trajec-
which the trajectory splits. The higher the Gini index,
tory, although this is only used as a criterion for further
the more different the velocity.
analysis, not within the summarisation process.

Velocity. Other way of detecting relevant points in trajec- Transform. There are other types of techniques that,
tories is the velocity. This metric is completely specific instead of using the trajectory points purely for seg-
to trajectories (as there is no temporality in line simplifi- mentation, they calculate N points that allow, using a
cation). It should not be confused with SED or TR, mathematical transformation, to reconstruct a continu-
which are distance metrics, although they are adjusted ous approximation of the original trajectory. An exam-
according to the time of the measurements. ple of these evaluation criteria is shown in Figure 6.
This metric uses velocity to summarise the trajectory, These techniques are very common when working
so it is no longer based solely on position in the plane. with one-dimensional (1D) time series. For example, in
This metric is more informed about the dynamics and audio or electrocardiogram data, but not so much in
context of the vehicle’s movement, allowing for more the trajectories field, which have several dimensions.
complex, even semantic, analyses. For example, unlike However, there is research that tries to apply this
distance, it can detect high-speed variability due to concept to trajectories. Rana et al.30,31 and Yuan
10 International Journal of Distributed Sensor Networks

a Bayesian Network, having also more types of possi-

ble movements.

Multi-criteria. Finally, many researchers choose to com-

bine several criteria to make a much more robust sys-
tem. This is indicative of the fact that most metrics
alone are not sufficient to model and detect all needs.
Figure 6. Transform criteria example. In Table 2, those will appear several times.
Researchers perform the joining of criteria in differ-
ent ways. Some perform criteria cascading. First, they
et al.27 propose to use Compressive Sensing, to obtain apply a segmentation algorithm, and then, on the seg-
the N-values with which to reconstruct the trajectory. ments found, a technique is applied that finds the most
This type of approximation is usually accompanied by representative points of that segment. Others perform
a trajectory filtering process to smooth the trajectory all the checks for each criterion simultaneously. Some,
and eliminate noise, making it more tractable. Yuan such as the TraClus,66 use a grouped metric evaluates
et al.27 use a particle filter, while Li et al.141 apply a distance and angle. This follows the Minimum
wavelet transform for each dimension of a maritime Description Length (MDL) principle, treating the
trajectory. Long et al.87 also test the wavelet transform problem as a cost minimisation. Several algorithms fol-
for direction preservation. The fast Fourier transform low this same principle. Others, such as online data
(FTT) can also be used as simplification criteria, as compression algorithm for trajectories (OLDCAT)26 or
Katsikouli et al.29 did to compare its solution. the proposals of Patroumpas et al.73,74 perform differ-
However, other investigations use splines, polyno- ent decoupled conditions in parallel, by means of a
mial lines that approximate the trajectory. These, unlike concatenation of comparisons. These techniques per-
the previous techniques, are lines that use the N dimen- form online compression to simplify the trajectory and
sions. Marino and Manic99 generate a continuous obtaining segments with semantic content.
spline that approximates the entire trajectory. Feldman Something similar is done by Siddique and Ban32
et al.34 also use splines, but generate several per trajec- with self-adaptive sampling (SAS) algorithm. It detects
tory, each one representing a segment of the trajectory. the dynamic variations of the vehicle (constant speed,
stopped, accelerate and decelerate) with an HMM and
use them to segment the trajectory (Vehicle Flow iden-
Probabilistic. All the previous techniques develop a spe- tification). It also uses a support vector machine (SVM)
cific analysis based on tangible, measurable, quantifi- classifier to detect if the car is stationary (so it is not a
able metrics. In contrast, the techniques of this trajectory, it does not move).
grouping perform an analysis with a more complex Others combine preserve criteria in several passes
algorithm, the results of which are probabilities. and with a combination of conditionals, such as Feng
These approaches mostly seek to classify each trajec- et al.115 with speed, distance and angle calculated
tory point as a specific type of movement. By grouping through the SED projection; or Gao and Shi94 with
in segments according to the type of movement of the angle and SED. Sánchez-Heres78 does something simi-
vehicle, more specific segments can be obtained for the lar to compress the straight lines but keep the turns.
type of solution desired. Depending on the type of The heuristics proposed above decide which points
movement, these segments can be further summarised to select. But, depending on the problem and its char-
to reduce the amount of trajectory points. acteristics, it may be necessary to slightly modify the
There are multiple approaches to classify the motion operation of the problem to adjust it. Therefore, several
of a vehicle.142 Siddique and Ban32,33 use Hidden of the proposals perform an additional analysis when
Markov Model (HMM) to classify trajectory points choosing the trajectory point to preserve. This analysis
according to whether the vehicle is stationary, varying is reflected in a weight, a numerical value that can ben-
acceleration or at constant speed. While Garcia et al.35 efit or harm the trajectory points, modifying the final
use an IMM estimation filter to obtain the type of air- decision criteria.
craft manoeuvre, being a solution implemented for air Li et al.108 to avoid selecting noisy trajectory points
traffic control.36 which introduces the weight concept to impair noisy
Zheng and colleagues127,128 make an analysis of the trajectory points. While Kulik et al.129 instead use
type of movement (vehicle/walking) of a person within weights according to the semantic content it wants to
the city, segmenting according to it. The analysis is per- preserve in its compression algorithm. Specifically, it
formed with different inference criteria. Feng and favours the preservation of major roads, while simplify-
Timmermans143 perform the same type of problem with ing non-major roads more.
Amigo et al. 11

Panagiotakis et al.53 and Pelekis et al.144 use a voting  Strategy combinations: these algorithms com-
criterion among several trajectories by distance between bine several of the previous strategies to make a
them to find the most representative subsegments of the more robust and intelligent approach.
whole set. Resheff76 does a version of maximum dis-
tance between segments radially, integrating also the
density of nearby points. Sequential. Trajectories are like chains, whose behaviour
depends on previous measurements. Therefore, they
should be measured considering the previous and ide-
Search strategy to process trajectory points
ally also the subsequent behaviour.
As already mentioned, all trajectory summarisation This category passes through all the trajectory points
algorithms need a strategy to process the raw trajectory in the simplest way, only once and in an orderly fashion
points. The strategy greatly influences the computa- through each of them. This strategy is used by research-
tional cost of the algorithm. The more basic ones only ers to analyse step by step whether the trajectory point
need to pass through each point once, while the more should be stored or not (decided by the score provided
advanced ones increase the order of complexity to a in section ‘Trajectory point evaluation criteria’), with-
large extent. According to the problem, different solu- out making double passes or comparisons by accessing
tions can be applied. The challenge is to find a trade-off previous measurements. This criterion allows to obtain
between computational time and quality of the solution. the best possible computation time, but the solution will
Throughout the literature, the following subcate- have a reduced quality, since it does not make enough
gories have been found that depend on the strategy used checks as others do.
for processing the trajectory points: It should be noted that this procedure is the one fol-
lowed with the proposals that perform external and
 Sequential: these algorithms follow the simplest
probabilistic analysis. It also models solutions com-
processing strategy, going through the trajectory
posed of many algorithms that need to extract values
points in order, analysing one after another.
initially, for which they make a first pass through all
 Window: these algorithms are based on the use
the measurements.
of windows that group several trajectory points,
Techniques that aim to find segments representing a
making the decision on the set of points. There
motion pattern usually employ this approach. This
are two main variants: a Sliding Window which
allows the movement along time to be analysed.
moves along the trajectory, or an Opening
Siddique and Ban32 achieve this through probabilistic
Window that gets bigger and bigger by adding
algorithms, while Wang26 does it with a multi-variable
new points to the evaluation.
analysis.
A variant of the latter is the use of estimation, which
is slightly different from the previous ones. In this case, Graph. The sequential solution allows a solution to be
the window checks whether a future estimate of the next found quickly, carrying out an analysis for each trajec-
points falls within the window. tory point individually. Even so, such a strategy is not
ideal for any problem, as its analysis does not have the
 Split: these algorithms are based on a segment capacity to contemplate previous situations that could
division strategy. An initial segment of the raw be positive for the summarisation.
trajectory is created, and it is checked if any tra- The opposite approach to sequential solutions is the
jectory point of the raw trajectory exceeds a one proposed by graph approaches. A graph models
threshold. If so, the segment will be split in two the entire search space, meaning all possible possibili-
and the process will continue iteratively. ties of the problem, to find its optimal solution. Graph-
 Merge: these algorithms are the opposite of the based strategies achieve an optimal approximation
split-based algorithms. Their objective is to start result but at a higher computational cost than others.
from a sequence of segments and merge two of In this problem of summarisation of trajectories, a
them consecutive using a threshold criterion. Directed Acyclic Graph (DAG) is used, where there is
 Graph: these algorithms are based on the genera- a predefined direction from the beginning to the end of
tion of a graph associated with the trajectory in the trajectory. The nodes of the graph represent each
which the nodes represent points of the trajectory trajectory point, while the vertices represent the sum-
and the edges represent the possible segments. marisation performed between two trajectory points,
discarding the intermediate ones (see Figure 7).
Once the graph is generated, the summarised trajec- Imai and Iri42 proposed a DAG by measuring all
tory is created by finding the best path within the distances between points of a line simplification, find-
graph. ing the optimal solution by minimising the error, but
12 International Journal of Distributed Sensor Networks

algorithm of Cao and Li41 solves this problem, being

able to generate the DAG with ISSED in an online and
incremental way. Wu et al.55 also explore real-time
graph solutions and compare them.
Another more specific type of solution is proposed
by Pulshashi et al.,54 this one is not so much focused on
compression, but on generating segments by eliminating
Figure 7. Graph strategy example. atypical trajectory points due to noise. To do this, the
DAG it generates introduces, for each trajectory point
the computational costs are too high for any solution. k new vertices, to the next k nodes. Moreover, with
It was improved according to the type of problem by these k additional paths per segment, it can develop the
Chan and Chin,43 making it feasible for some non- algorithm in real time.
priority problem.
This type of solutions can also be used to find the
optimal solution without using distances as a criterion. Window. Knowing that optimal algorithms are not a
For example, Long et al.60 implement a DAG that uses feasible solution for virtually any solution, many
angular deflection to achieve the best possible summar- researchers opted for other heuristic paths. These seek
isation given a maximum deflection. Long et al.60 also to find a path by performing without generating a
propose a solution (SP, also known as DPTS) in which DAG, but by traversing the trajectory in other ways.
it makes a graph applied to an angle heuristic, so that One of these uses a window that limits the search space
it finds the optimal solution. to a sequence containing a few trajectory points. This
Optimal solutions are desired in any problem, but in causes the preservation criteria checks to be performed
few problems can they be computed because of their only between such measurements, thus making the
high running cost. This makes it impossible to use these checks much fewer.
algorithms in real-time solutions in any case. It should be noted that the sequential concept can
Because of this, some researchers sought to find solu- be viewed as a window of size 1. This concept of evalu-
tions using graphs capable of achieving suboptimal solu- ating all consecutive points makes clear a linkage to be
tions, but whose computation time was considerably able to run the algorithm in near real time, having to
less. Kolesnikov and colleagues49,65,145 were the first store a few measurements to extract results. There are
researcher to introduce the concept of Reduced Search two types of approaches using the window concept to
Dynamic Programming (RSDP) on the simplification traverse the points.
problem. In the search for a balance between optimal
solutions, called Full Search Dynamic Programming Sliding window. For a particular trajectory point of
(FSDP) with a reduced computation time. These solu- the trajectory, N of the neighbours of that point are
tions, instead of generating the entire search space, base selected and the calculations are performed with them.
their operation on the limitation of the search space to The placement of the window with respect to the point
be computed from each vertex node. can vary according to the algorithm: placing it in an
Initially, Kolesnikov and Fränti145 perform the pro- intermediate way and taking points before and after or
cess in a decoupled fashion, initially using a DP to have taking only points on one side.
a reference solution, and exploring alternative paths for Once the operation of the window is established, the
each node. He then improved his solution by calculat- behaviour of the window depends on the algorithm that
ing both the reference solution and the alternatives at implements it. The trivial algorithm of Tobler130,131 can
the same instant.49 It used accumulated heuristics as be seen as a window of fixed size. The same is true for
the already explained ISE. Since then, several research- the Pulshashi et al.54 with graphs, which generates k
ers have opted to follow its path applied to trajectories, vertices for each point, where k is the size of the
seeking to find suboptimal or hybrid solutions, but fea- window.
sible for real-time use. Keogh et al.146 proposed a sliding window imple-
Chen et al.44 were one of the first to apply it to tra- mentation for time series. It checked the error in dis-
jectories. His MRPA algorithm uses the accumulation tance of the segment formed by the window, with
of the SED distance metric, ISSED, to find the shortest respect to the intermediate trajectory points. If they
path. This solution follows the performance of DAGs, exceed a threshold, it splits, otherwise it shifts the
but optimises it computationally using only two queues whole window, discarding them from the summarised
with priority. Likewise, this solution requires the com- trajectory.
putation of all distances before finding the solution, Yan et al.112 instead use the window to find and clas-
which still penalises its real-time execution. DOTS, the sify fragments of the segment according to the type of
Amigo et al. 13

movement. Thanks to the window you can apply it in Split. Another strategy to go through all the points
real time. quickly is the Split method. This method is based
The solution of Muckell et al.,100,101 SQUISH, uses a on Dynamic Programming, generating a recursive
window of N measurements. On these N measurements, way of dividing the global problem, and solving it in
it chooses the worst one and eliminates it, inserting the smaller pieces. Also known as divide and conquer, or
next one and so on until the trajectory is finished. top-down, the objective of this strategy is to split the
Marino and Manic99 use the Window to quickly cal- trajectory by the most relevant point according to a cri-
culate the direction correlation of the trajectory points. terion. With this process, two subsegments are gener-
Finally, Gao and Shi94 use a window on which you ated, one from start to the selected point, and the other
apply several criteria at the same time. from this point to the end. These two segments, in turn,
apply the same problem again, generating the recur-
sion. In this way, the trajectory is segmented until it
Opening window. Alternatively, the window can be
reaches a previously established limit. This approach
viewed as a point buffer, where as long as a criterion is
gives fast results, but its way of traversing the trajectory
active, trajectory points continue to be entered. When
points prevents it from working in real time since it
the criterion is no longer met, the window starts again.
needs to process the whole trajectory.
Depending on the implementation, either at the last
The most known algorithm of summarisation,
point of the window, which is preserved in the summar-
Douglas and Peucker114 (DP), uses this strategy with a
isation, or at the last current point of the summarised
PED distance threshold. Thanks to it, multiple
trajectory.
researchers have tried to improve it. For example,
The first reference found in the literature that
Hershberger and Snoeyink116 proposed a variant with
applies an opening window is the proposal by Shatkay
the same result but faster, called DP-hull.
and Zdonik,147 applied over time series. Although the
The algorithm Scan, Pick and Move (SPM)120 is a
main author of this approach is Meratnia and De By,3
variant similar to DP, but instead of generating two
which introduced the name of this concept, opening
segments on which it is necessary to reapply the algo-
window. In addition, Meratnia and De By3 differen-
rithm recursively, it keeps fixed the first of the two seg-
tiated between two variants, whether to stay with the
ments generated, applying recursion only on the second
point that exceeds the window, or the point just before
segment. This provides on long trajectories a faster
(Normal OPW), where the window has not been
solution, but at the same time the result will be worse
exceeded (Before OPW). Another variant to find the
and does not guarantee a maximum error.
specific point is that of Meng et al.,70 which after
Like all strategies, the split-based can also imple-
applying the window with a criterion, uses a cumulative
ment any type or combination of point preserving cri-
SED distance to find the exact point to store in the
teria. For example, Liu et al.125 use it to minimise the
summarisation, and continues from there.
area of the MBRs that encompass the trajectory points
Many researchers followed this type of trajectory
of a trajectory.
approaches, for example, Lee et al.66 and Sheng et al.79
The segments are then generated with their multi-
criteria metrics, until it exceeds a threshold. The same Merge. The opposite way to Split can also be applied
is done by Liu et al.89 using speed as a criterion. on trajectory data. Instead of starting from the general
Direction estimation algorithms are also related to problem and going to multiple specific problems and
opening windows. The algorithms fix an orientation finally joining the solutions, the merge algorithms start
and all points that fall within this window are not from multiple specific problems and arrive at the solu-
selected for the summarised trajectory. As this enters tion of the general problem. This strategy is also consid-
algorithms such as Reumann and Witkam77 establish- ered as Dynamic Programming, it starts with segments
ing two infinite parallel lines, Opheim102,137 with a of few trajectory points and the algorithmic process
delimited area or even Dead Reckoning81 and its var- oversees finding which pairs of segments can be joint
iants Connection-Preserving Dead Reckoning together in bigger ones.
(CDR),50,148 which also recalculate the intermediate This process is also known as Bottom-up, or elimi-
points of the window at each iteration, in case it moves nation, because, when joining segments together, there
far enough away from the current point vector. is a trajectory point that is discarded. Unlike the Split
Kolesnikov65 does something similar, setting an area strategy, this one can be applied in real time with sev-
where the point can be as a prediction, and the follow- eral nuances.
ing points must be within it. In addition, thanks to the Pikaz and Dinstein109 performed the first appear-
window, it performs the combination of the areas of ance of Bottom-up in this type of problems, applied to
past points, obtaining a more accurate estimate. polygon approximation. Hunter and McIntosh149 also
14 International Journal of Distributed Sensor Networks

Table 3. Special approaches table.

Special approach Techniques

Multiple trajectory compression Similarity,113 TrajStore,151 Representativeness,53 NaTS144

Lossless trajectory compression PRESS,25 COMPRESS,96 CoTracks,90 Trajic,52 TrajStore,151 IFC,152 Lovell,12,153 LWZ154
Network road constrained PoI and PoIE,61 Nonmaterial,155 VTracer,58 TSHL,63 OGPC and OSPC,47 Opheim-
improved,95, RSLC and TSLC,96 CFF,64 MMTC-offline and MMTC-online,48 FFDP and
FFUS,97 GS,108 IC-MBR,125 SUTC,89 INCM,118 STTrace and Thresholds,75 ESTC-EDP,119
STC,22 BTC and HTC,25 STMaker,103,104 SNDSC,122 CLEAN57
Semantic segments PoI and PoIE,61 TS,127,128 2stage-pls,126 IMM,35–37 OGPC and OSPC,47 FFDP and FFUS,97
S-DMin and SE-DMin,129 ATS,117 SMoT,71 CB-SMoT,21 Patroumpas,73,74 STC,22 RGRASP-
SemTS,24 BTC and HTC,25 STMaker,103,104 SELF,121 SetraStream,82 HESAVE and SNDSC,122
SPD84

applied it in time series. Visvalingam and Whyatt111 the literature. Keogh et al.4 proposed that union
apply this concept also with an area criterion. It starts between a window and merge strategy, where the slid-
with trios of trajectory points and eliminates the inter- ing window incorporates points until it fills a buffer,
mediate point of each trio when it has the least area of and merge empties the buffer, selecting the most rele-
the whole trajectory. In this way, the trio that has been vant points.
eliminated disappears, and the others are modified, Another interesting mix is the one performed by Liu
using the closest point that has not been eliminated. et al.,125 where it starts from segments of fixed size
In trajectories it is possible to highlight, SQUISH100 (number of trajectory points that make up a segment).
and its improved version SQUISH-E101 which are two On them, apply Split or Merge according to the size of
of the recent algorithms with the greatest impact in the the segments, until the MBR area is minimised. The
literature. Both use a merge strategy, mixed with a slid- already mentioned SQUISH-E101 and STTrace75 are
ing window to minimise the SED distance. The algo- also examples of a combination of merge strategy with
rithm is highly configurable and can be set to minimise a window that limits the amount of information to be
the error while maintaining a specific compression processed.
ratio, or the opposite, to maximise the compression Other approaches do decoupled cascading during
ratio while keeping the error below a specific value. execution. These detect, for example, when to store a
Another similar solution with merge strategy and point, so that a later analysis selects which of the points
window is the one performed by the algorithm to store. This is the case of Meng et al.,70 who first
STTrace.75 It first fills a buffer and then removes the applies an opening window until an SED distance
one with the worst SED. Li et al.108 also execute a threshold is exceeded, but once the threshold is trig-
merge strategy, eliminating the points with the highest gered, finds the point to be preserved in Split form,
weight, calculated in a multi-criteria way. with a different metric.
The same happens in the Generic Remote Trajectory
Strategy combinations. Finally, as is also the case in tra- Simplication (GRTS) algorithms proposed by Lange
jectory point preserve criterion, strategies can be com- et al.148 It uses the CDR algorithm to detect the
bined with each other to realise solutions with different moments to send points, using an opening window
objectives. In Table 2, those will also appear several strategy. Once out of the threshold, apply another,
times. more accurate algorithm to compress the fragment. It
The need for graph algorithms to limit their compu- can even run high computational time algorithms
tation time in order to achieve a suboptimal but fast online, by reducing the number of measurements it
computational solution was discussed earlier. This must process. GRTSOpt uses an optimal algorithm
effect is repeated in other types of strategies, already such as Imai and Iri,42 whereas GRTSSec uses a heuris-
more efficient than a hybrid graph, achieving an imple- tic solution OPW-TR.3
mentable solution in IoT boards or low-cost beacons. Finally, the algorithms that perform the cascade in a
The opposite effect can also be sought, to find a more totally decoupled way. They first make a pass with one
specific solution with a more refined result. strategy and with the results, apply another or other
Some apply the different strategies at the same time, strategies with which to achieve summarisation. A typi-
supporting each other. Sliding Window And Bottom- cal cascade approach is by first carrying out a sequen-
up (SWAB) is one of the oldest combinations found in tial run that extracts values from the trajectory, and
Amigo et al. 15

trajectory points, it allows a more detailed analysis of

the target’s behaviour. Although the concept comes
from there, some researchers use it to refine the seg-
ments without techniques that work with probabilities.
Garcia et al.35 use an estimation filter called IMM,
the result of which lets you know what type of motion
the target is performing. If set correctly for a particular
type of vehicle, this filter can estimate by probability
the type of motion the vehicle is making. The forward
pass is responsible for detecting the beginnings of a
motion segment, while the backward passes detect the
ends. Their union allows to generate informed segments
categorising straight, turning, stationary and so on
Figure 8. ATS117 algorithm illustration. Multi-criteria by movement.
cascade. Liu et al.98 perform a double pass in time series,
where the first Feasible Space Window (FSW) finds the
segments following a distance criterion and a sliding
then, using these, other algorithm is applied to select window strategy. Then, the backward pass, Stepwise
the measurements. FSW, retouches the position of the segments reducing
Hansuddhisuntorn and Horanont17 use this the representation error of the initial segments of the
approach to reduce the number of trajectory points to FSW. This approach allows to make a final pass when
be introduced into an algorithm, so that it is less com- the trajectory is completed, so that in real time the seg-
putationally expensive. First, it makes a sequential ments are suboptimal, but when finishing and storing
strategy to choose only relevant points in velocity or the results, they are improved with the backward seg-
angle, and those it inserts into the algorithm Top- menting process.
Down Time Ratio (TD-TR).3 The same objective has Kolesnikov65 with its Scan-
Lee et al.150 make two forward passes. First, it does Along Trajectory Approximation (SATA) algorithm,
a coarse partitioning of trajectories separately (using explained above. By accumulating the areas found by
TraClus). Then, with the trajectories potentially out- means of an opening window, it guarantees a minimum
liers, it makes another more specific segmentation by of error. However, this single pass can be improved by
pairs of distinct trajectory segments. performing another pass in the opposite direction,
Another example of the use of velocity in summari- applying the same window now in the other direction.
sation is proposed by Lin et al.117 As is shown in Figure Tang et al.80 and Etemad et al.91 use a double-pass
8, its algorithm uses the Gini index on the velocity val- approximation in a different way. For certain trajec-
ues to detect the points at which the trajectory splits. tory points, it applies the double-pass concept along a
The higher the Gini index, the more different is the sliding window. With this it checks the velocity (Tang)
velocity. Then, with the velocity values, find the error or the direction (Etemad) of several points before and
threshold to use from PED in a DP to compress for after to improve the detection. Tang later applies an
subsegment. improved DP for segmentation.
DOTS-CASCADE41 applies the DOTS algorithm N Finally, as summary of this section, Table 2 indi-
times by parallelising the network computations, so cates where each of the algorithms studied in the litera-
that it can run in real time. A similar approach follows ture are located accordingly to the categories proposed.
Siddique and Ban32 using first HMM to locate the Notice that some of them combine several criteria to
important points and then SVM for stops. Other algo- make more robust solutions, and for that reason, they
rithms are based on applying several passes (each pass appear in several places in the table.
following a procedure different from the rest) on the
trajectory points to acquire knowledge that cannot be
Special approaches
obtained by means of a single pass.
Within this last category, a common operation can Within the trajectory summarisation literature
be identified: algorithms that perform passes of the reviewed, in addition to all the groupings above, some
same algorithm in both directions of the trajectory. One special approaches to the problem have been found.
pass from the beginning to the end, called forward, and These still aim at reducing the dimensionality of the
another in the opposite direction, called backward. This trajectory, but they have an additional motivation that
type of strategy is common in probability algorithms, makes their implementation peculiar. The most rele-
but since there are probabilistic criteria for preserving vant of these are explained as follows:
16 International Journal of Distributed Sensor Networks

 Multi-Trajectory Compression (MTC): instead Birnbaum et al.’s113 technique generates segments

of compressing each trajectory individually, they for each trajectory using a lossy STC algorithm, and on
try to take advantage of the knowledge coming the other hand tries to represent the trajectory with sub-
from the existence of multiple trajectories in rela- segments of other trajectories already stored. The form
tion to each other. that minimises the error will be the one that stores, if
 Lossless compression: these algorithms stand out the new segments obtained with the compression STC,
for generating the summarised trajectory without or it will use the already existing ones in the MTC. It
any loss of information, being possible to recover also minimises the replicated information, namely, by
the raw trajectory from the summarised one. storing the start and end time of each segment, since all
 Road networks constrain: these algorithms stand the intermediate positions can be interpolated from the
out for being specially designed to be used in base segment. In addition, the successive points store
conjunction with context information from the only the gaps between measurements and time, not the
road network from which the information has complete value, to reduce space. This facet is usually
been taken. related to lossless compression, which is more oriented
 Semantic approach: these algorithms are notable to databases.
for including and using semantic information, Zheng and colleagues28,156 propose a framework
extracted from the applicable problem, when that first eliminates the redundancy of multiple trajec-
generating the summarised trajectory. tories, keeping only the priority segments, and then
compresses each segment separately. To do so, it uses
This categorisation is shown in Table 3 and also at similarity metrics, which compares all trajectory points
website.15 of both segments with each other.
Likewise, any trajectory subsegment clustering tech-
nique, such as TraClus,66,157 could be applied to this
MTC criterion. Panagiotakis et al.53 and Pelekis et al.144 use
a voting criterion among several trajectories by distance
Trajectories usually have a predetermined destination, between them to find the most representative subseg-
and, if such a trajectory takes place with a certain fre- ments of the whole set.
quency, it usually has the same path that is followed to
get from one point to another. These trajectories predo-
minate in urban land navigation, where almost all vehi- Lossless compression
cle movements are on roads or at least dirt roads. In One of the most common characteristics of summarisa-
places where movement is not restricted, such as mari- tion algorithms is the loss of information suffered with
time navigation or air traffic, and although less so, a the compression of trajectory points. This means that
series of prefixed paths are also usually followed on the algorithms manage to generate segments containing
long routes. new information based on the original measurements,
There are several approaches that seek to exploit this although the generated segments do not represent the
redundancy of trajectories on the same path to maxi- original information. This approach is the most com-
mise the summarisation and compression of a trajec- mon and is called lossy.
tory. The objective is, instead of compressing each However, there is another approach in which the
trajectory individually Single trajectory compression information is compacted without degrading it. This
(STC), like all previous techniques, to try to exploit the approach, called lossless, makes it possible to preserve
knowledge of existing trajectories with similar shape the original information while occupying as few bytes
and dynamics. In doing so, solutions can reduce dimen- as possible. In addition, it must ensure that all the pre-
sionality abruptly, going from hundreds of redundant cision of the raw trajectory is recoverable. This type of
trajectories in shape, to storing only one that represents approach usually gives little importance to the sum-
all of them. marisation part and focuses entirely on data reduction,
These approaches are usually more related to a com- although some techniques do exploit the characteristics
plete framework for a posteriori analysis, or clustering of the trajectories.
techniques, as they require multiple refined trajectories The simplest example to understand the difference
ready to be processed and compared with each other. between lossless and lossy compression is to switch to
TrajStore151 was the first approach to do MTC. It another domain. Images can be lossy compressed into
groups virtually identical trajectories by clustering, JPEG files, which achieve a very high compression
comparing the trajectories with each other using simi- ratio, but generate artefacts in the image that do not
larity metrics. When it identifies a cluster, it uses one of exist. However, if compressed in PNG files, lossless
the trajectories in the cluster as a representative of that compression is achieved in each pixel of the image,
group, saving the storage of the N trajectories. although with a lower compression ratio.
Amigo et al. 17

Lossless bases its operation on structuring the data Therefore, with the aim of making applications for a
in a more efficient way, so that the information is com- non-expert public, the transformation of this numerical
pressed. In addition, it looks for redundant patterns in information into tangible, legible, explainable knowl-
the information to save space. Transforming the data edge is a very well-studied and necessary problem for
to such a structure requires extra computation time, any application. This knowledge is obtained by means
and the same for decompressing it again to obtain the of artificial intelligence techniques, capable of analysing
original path. These techniques have disadvantages, the behaviour of thousands of trajectories and knowing
such as making it difficult to access the trajectories with how to differentiate a specific aspect of each one of
such intermediate computation, but they have advan- them.
tages, such as storing the same original trajectory at a This review considers the semantic content in trajec-
lower space cost, useful for long-lived databases. tories in two ways: first, by generating self-explanatory
Lossless compression is a commonly practised term segments, according to the type of movement the vehi-
in computing. In fact, the Consultative Committee for cle performs. Or second, by generating additional infor-
Space Data Systems (CCSDS)158 proposes recommen- mation related to the geographical context through
dations and sets the standard for how a lossless com- where the trajectory runs.
pression algorithm should work. OLDCAT,26 Patroumpas et al.,73,74 Siddique and
There are two types of lossless algorithms: the gen- Ban32,33 or Garcia and colleagues35–37 are authors of
eric ones, applicable to any type of computer file with a approaches that detect the change of trajectory motion
decent compression ratio. The other approach is to type. This implies that they are able to identify the type
develop a specific solution for the data to compress. of motion between two points of change, categorising
Like in the mentioned example with image data, it is segments with a particular type of motion.
possible to develop a specific proposal to deal with tra- Previously some algorithms were mentioned aiming
jectory data. Below, solutions in the latter category are to find the start-stops points of the trajectory and com-
presented. pressing with it. These approximations are semantic
Hatanaka159 eliminates redundant information, stor- content. Zheng et al.84 apply clustering of individual
ing only the gaps between measurements, reducing trajectory points to detect stay points. Alvares et al.71
information by up to 80%. Song et al.25 propose did something similar, but the detection is done seman-
PRESS, a map-matching framework with both trajec- tically, by placing an area of interest to monitor. When
tory summarisation modes: lossless and lossy. Then, more than one time threshold is found within the area,
Han et al.96 proposed COMPRESS that starts from the it is a stop. Between stops, there is movement. It detects
base proposed by Song and improves it different fields. interesting trajectory points on ships from the angle of
TrajStore151 compresses the trajectory by clustering rotation between measurements using clustering. They
and saving a representative of each cluster. Later, found this algorithm useful to identify fishing spots.
Trajic52 claims to have outperform TrajStore approach. Tamilmani and Stefanakis121 use the Semantically
Trajic’s paper also develops a different lossy solution, Enriched Line simpliFication (SELF) structure to store
based only on bits encoding. semantic content, aside from position and time. It spe-
There are even proposals to combine both forms of cially compresses the semantic content by angle and
compression as Balzano and Del Sorbo90 does: first a velocity, allowing it to be interpolated if necessary.
lossy compression is perform. Then, with the selected Richter et al.22 propose a compression in road net-
trajectory points, a lossless compression, so they occupy works using map-matching, but, unlike the previous
as little space as possible. ones, it does not store the positions where it is located,
Lovell12,153 has recently made several approaches but stores the name of the streets it travels, a more
seeking to exploit kinematic values to perform lossless understandable way for the human. With this informa-
compression. In addition, the paper provides a clear tion, it is still possible to decompress and find the real
overview of the lossless trajectory compression status. trajectory, together with the time. Su et al.103,104 take it
one step further, summarising the trajectory in natural
language: its crossing points, average speed in each seg-
Semantic segments ment and so on.
Many approaches in the literature, especially recent
ones, aim at generating segments that represent con-
crete information. These are called semantic knowledge. Road network constrained compression
The trajectories are compiled by GPS sensors, giving a There is another type of approximations when the tra-
position over time, moving over the Earth. These, in jectory occurs in a network of roads that limit the tra-
computational form, are tuples of numbers with several jectory performance. Here the problem of GPS noise is
decimals, which a human being is not able to under- accentuated, appearing noisy trajectory points that go
stand, at least, without a previous study of the problem. out of the trajectory, being necessary a previous
18 International Journal of Distributed Sensor Networks

iteration that adjusts these points to the corresponding and its significance in the most recent studies will be
road. Subsequently, with the points already adjusted to explored. In particular, the following characteristics
the road, the segmentation/compression algorithms use have been identified:
the context of the road, mainly the intersections, to
approximate the trajectory.  The ability to segment in real time as soon as the
Moreover, with the concept of roads, it is not neces- data are available. There are two possibilities,
sary to represent and store the shortest path between batch mode implies that the complete trajectory
two points, but it can be inferred a posteriori if needed, data are available after the entire trajectory has
since the possibilities are reduced. This problem could been traversed. The online mode means that the
be done with Naı̈ve solutions, combining the two algo- data are available in real time, as each measure-
rithms: Kellaris et al.48 propose to do compression and ment is taken, the data are passed to the algo-
then map-matching. Or applying map-matching first, rithm for processing.
with the points on the road, compress, and then map-  The type of input data implies the level of trajec-
matching to the road but taking up less. tory information used. That is, the time compo-
Non-material155 was the first to combine trajectory nent of the trajectory is taken into account, or
compression and trajectory map-matching. His solu- only the shape of the trajectory.
tion separates the spatial trajectory, which can be extra-  Finally, due to the flexibility and complexity of
polated from roads, from the temporal component, these algorithms, it is explored if it is necessary
which belongs to each track. The spatial compression to adjust their parameters accordingly to the
stores the crossing intersections and the temporal gaps problem and the type of trajectories to be
between intersection pairs. summarised.
Map matched trajectory compression (MMTC)48
uses subtrajectories through fewer intersections to Like the other classifications, all algorithms are also
replace parts of the original trajectory. Some specific categorised according to these features. Available at
evaluation functions are introduced during the com- website.15
pression to guarantee the similarity between the com-
pressed trajectory and the original one. The compressed Real-time operating
trajectory consists of fewer intersections; thus, the stor-
age cost is reduced. There were two main ways of approaching an algo-
Gotsman and Kanza47 proposed several ways to per- rithm to process data, depending on the availability of
form compression in road networks. Using graphs, it the data, it is possible to work in real time as new data
finds the shortest path (highest compression), knowing appears or offline after all the data have appeared.
that it can then redo the path (since the path can only Working with trajectory data, an offline or batch algo-
pass through the available roads) looking for the short- rithm, because of its way of processing data, requires
est path between both compression points. It proposes all the trajectory points from the beginning of the exe-
optimal and even online solutions. cution. Having all the information from the beginning
Li et al.108 work offline, taking into account the confi- allows them to provide potentially optimal solutions to
dence in the GPS measurement, so it eliminates outliers. the problem. Meanwhile, a trajectory point buffer can
It applies the compressor first and then does map-match- operate in real time, delivering results as more measure-
ing, so it does not link the two phases. Something similar ments arrive. This needs to provide results in real time
does Cui et al.63 which first segment with angle and means that they cannot claim to find the optimal solu-
tion. They must perform a trade-off to obtain a good
length and then fit the segments to the road.
Liu et al.89 first do map-matching and then check if solution within a computation time that allows process-
ing measurements faster than the time in which new
the speed is adequate and applies lossless compression.
measurements arrive.
Popa et al.118 propose another compression method
Both approaches are useful in a problem as broad as
with deterministic error bounds and an error measure
trajectory analysis, which has so many possible uses.
for in-network trajectories. It assumes that the noise
For example, in a big data environment, where all the
path has been cleared, and all the points are already on
available information from millions of trajectories is
the road.
available, a batch solution will be preferred, potentially
with better results. However, there are use cases where
the solution must use an online algorithm. For example,
Other characteristics
when monitoring a target using a mobile device, it is
As mentioned previously, a literature review with a necessary to process the detections and obtain results in
good number of papers will detect several common seconds with which to make informed decisions. There
characteristics. In this section, they will be introduced is also a middle ground, and it should be considered,
Amigo et al. 19

when the entire trajectory is available, but the solution trajectory, since it is as or more important than the rest
has to be delivered with a relatively low delay. of the variables.
The original line simplification algorithms did not The current trend is to try to add another type of
require immediacy in the results to be obtained, as they semantic knowledge to generate representative seg-
were static data with hardly any online use cases. In ments, both context of the segment with respect to the
addition, the sensors were less accurate and the amount rest of the trajectory, and of the environment through
of information per trajectory was limited. Therefore, which it moves, or the difference with respect to other
most of the algorithms in the literature were batch. vehicles travelling through the same area.
Starting in the 2000s, trajectory summarisation algo- An important factor for the correct development of
rithms began to be developed. Initially, they start from trajectory simplification algorithms in the future is that
line simplification algorithms, but due to the technolo- they correctly take time into account. Algorithms that
gical advance, other uses have been generated and the do not consider time and are based solely on shape are
literature has covered them. Nowadays, there are many outdated for most of the analyses to be performed.
more and more accurate ones. The use cases of these They are clearly inferior for any trajectory problem.
technologies are looking for online solutions that pro-
vide segments that represent additional information,
beyond the mere compressed line. The accuracy and Need to adjust parameters
redundancy in a trajectory are so high that it is not so Most of the summarisation algorithms have parameters
much necessary to maintain it, and have it taken up less that can be adjusted to obtain a correct performance.
space on the device, as it is to extract useful informa- For this, an analysis of the type of trajectories to be
tion for various contexts. simplified, the characteristics of the algorithm and each
This needs to have solutions as soon as possible has parameter is necessary before running it.
led to a recent trend where researchers try to exploit all Most of the works presented above perform the
the computational characteristics of the devices where adjustment of parameters by hand by trial and error,
they implement their solutions. Currently, processors making modifications of the parameters until an accep-
have multiple cores that allow parallelisation of compu- table solution is found that represents a local optimum.
tation. Others even have dedicated components for Some examples are Zhao and Shi161 or Amigo et al.162
high-performance tasks, such as artificial Intelligence which perform an empirical manual analysis of the
(AI)-specific cores or graphics processing unit (GPU) threshold fit to find the best segmentation.
cards with thousands of cores. This parallelisation However, some algorithms do not perform a manual
should be in the design of the online and offline com- study but have an automatic and unsupervised adjust-
pression algorithms of the future and can accelerate ment of the parameters. Liu et al.20 automatically fit it
very expensive implementations by many orders of inside a DP, for the maritime context. Something simi-
magnitude. Some examples of the parallelisation are lar can be observed in the study of Zhang et al.163
the Patroumpas et al.73,74 approaches, graph speed-up Many parameters to make it work properly. Some
by Deng et al.,59 Feldman et al.34 or Huang et al.,160 or researchers use multi-objective evolutionary techniques
the spin line detection of Feldman et al.34 to fit the parameters within their problem.164,165
In Shuang et al.’s166 solution, an automatic speed
threshold is calculated within segments for anomaly
Input data detection. While Wei et al.’s107 solution uses statistical
Related to the evolution of the techniques over the theory is applied to determine thresholds for course var-
years is the evolution of the type of data used by the iation and speed variation.
summarisation algorithms. Most of the techniques ini- Soares et al.23 use MDL to avoid the use of thresh-
tially developed were designed and tested by their olds. This solution selects N points randomly as repre-
authors for two different problems: time series, that is, sentative points, similarly to a clustering approach, and
a variable over time, or line simplification, that is, posi- automatically adjusts itself by means of the cost func-
tion only, without time. These were used as the basis tion formulated with MDL.
for this new branch of research, which relies on the
union of position and time.
Summarisation evaluation
There is a trend where summarisation algorithms
are moving away from using only the trajectory form, Since there are so many algorithms and approaches
making solutions that could also work for line simplifi- that summarise trajectories, it important to analyse the
cation problems, and are incorporating all kinds of performance of each algorithm in the different prob-
additional variables. The first step was to introduce the lems, being capable of comparing them to select the
time component in the calculation to summarise the most suitable for each case.
20 International Journal of Distributed Sensor Networks

Some of the articles reviewed do not make any eva- shows the comparison provided in the paper that intro-
luation of the performance of the algorithm, especially duces each algorithm.
when summarisation is merely a step towards another This review provides an initial overview of multiple
end purpose. This practice is undesirable. It is essential existing approaches in different categories, facilitating
to always be informed whether the segments met the the decision of with which algorithms a future work
qualitative and quantitative needs of the next step in should be compared. Also, a set of highlighted algo-
order to guarantee the results will be achieved. This rithms is also provided in the introduction according to
work analyses which metrics they use to evaluate its the type of solution desired.
work, against which algorithms they compare its solu-
tion and which type of trajectories data is used as input.
Data sets used
Metrics A proper evaluation of a summarisation algorithm
requires that others can evaluate their algorithms on
For the evaluation, many evaluation metrics have been the same set of trajectories and benchmark it against
used in the reviewed literature that check how well a them.
compressed trajectory (or the generated segments) Throughout the literature, there is a clear predomi-
approximates the real trajectory. Those metrics can be nance of trajectories in cities. This is logical, as this is
divided into the following categories: the domain where the majority of travel occurs for
humans. This also leads to the existence of many spe-
 Summarisation meta-information: metrics that cific algorithms for such trajectories. Some algorithms
compare the raw and summarised trajectories use real experiments with their own data, but several
regardless of the trajectory factor. Metrics such public data sets stand out in the literature.
as compression ratio or bytes reduction are GeoLife167 is a data set that is developed in a city,
metrics in this category. with trajectories categorised according to the motion
 Algorithm computational complexity: these method: vehicles, bicycles or people walking. Data sets
metrics seek to measure the computational effi-
of taxis or trucks moving through a city are also com-
ciency of each algorithm, by checking the execu-
mon. Data sets of taxis28 or trucks62 moving through
tion time or memory consumption.
road networks are also common.
 Error metrics: metrics that calculate an error
Other researchers use ship data. AIS technology is a
residual between the actual trajectory and the
simple approach to data set generation, as it is an open
summarised trajectory. The most common way
standard by which ships are required to communicate
is to adopt the trajectory point preserve criterion
their position. There are multiple data sets that collect
used alongside a statistic values (average, maxi-
information from AIS detections in specific areas.168
mum, etc.).
These data also incorporate contextual information
 Similarity metrics: using specific metrics that
such as ship type or destination, allowing for easy post-
compare two trajectories (or segments) to see
summarisation analysis.126,169 There are other types of
how different they are. Some works use this
trajectories that are less used but noteworthy. Among
metrics comparing the real trajectory and the
them, there are some hurricanes data sets,91 with a
summarised one.
more limited amount of information, but with very dif-
ferent behaviours from the rest of the trajectories.
Comparison to other algorithms Finally, the trajectories of tracking animals: such as
deers60 or bats.68
This article also explores how works benchmark their
results against others in the literature. This is the only
way to demonstrate if an algorithm performs well or Conclusion
not for a specific problem (the problem described on
each paper). In this article, an overall review of trajectory summari-
Most papers do not compare themselves against sation algorithms was provided, merging both compres-
other algorithms in the literature, showing only the sion and segmentation concepts in a same perspective.
results they achieve. This lack of knowledge of the This type of approach has not been explored in any of
existing literature on the problem means that many the existing surveys available in the literature.
researchers consider its approach innovative when it is The concept of summarisation and its application in
possible that previous work did it earlier. Moreover, it different use cases has been introduced. These have two
is possible that another algorithm in the literature can main motivations: the need to shrink or simplify the
perform better for the same problem. Column trajectory to lighten the workload of further algorithms
‘Comparison to other algorithms’ on the website15 or to extract more knowledge from the trajectory.
Amigo et al. 21

Two main categories were found to group the more 5. Meratnia N and De By RA. A new perspective on trajec-
than 160 algorithms found in the literature: the search tory compression techniques. In: Proceedings of the
strategy for the trajectory points and the criteria to ISPRS commission II and IV, working groups II/5, II/6,
decide which of them to preserve in the summarisation. IV/1 and IV/2 joint workshop on spatial, temporal and
multi-dimensional data modelling and analysis, Quebec
Distance as a preserve criterion is the most used, while
City, QC, Canada, 2–3 October 2003, 8 pp. Nice: Inter-
the strategy algorithms that have been found are all
national Society for Photogrammetry and Remote Sen-
balanced. sing (ISPRS).
Throughout the study, certain special trends were 6. Anagnostopoulos A, Vlachos M and Hadjieleftheriou M,
discovered, which were analysed in detail. MTC, et al. Global distance-based segmentation of trajectories.
Lossless or Road network constrained solutions focus In: Proceedings of the 12th ACM SIGKDD international
on higher compression, while semantic approaches aim conference on knowledge discovery and data mining
at generating additional context for the segments. (KDD’06), Philadelphia, PA, 20–23 August 2006, p.34.
Finally, additional classifications are made on all the New York: ACM Press.
analysed algorithms, from other points of view. For 7. Fu T-C. A review on time series data mining. Eng Appl
Artif Intel 2011; 24(1): 164–181.
each one, the evolution of the algorithms over the years
8. Lovrić M, Milanović M and Stamenković M. Algorith-
is explained and those trends that should be further
mic methods for segmentation of time series: an over-
explored in future works are highlighted. All of these view. J Contemp Econ Bus Iss 2014; 1: 31–53.
categorisations, which allow algorithms to be compared 9. Feng Z and Zhu Y. A survey on trajectory data mining:
with each other, are available on the website.15 techniques and applications. IEEE Access 2016; 4:
2056–2067.
10. Sun P, Xia S, Yuan G, et al. An overview of moving
Declaration of conflicting interests
object trajectory compression algorithms. Math Probl
The author(s) declared no potential conflicts of interest with Eng 2016; 2016: 6587309.
respect to the research, authorship and/or publication of this 11. Da Silva CL, Petry LM and Bogorny V. A survey and
article. comparison of trajectory classification methods. In: Pro-
ceedings of the 2019 8th Brazilian conference on intelligent
Funding systems (BRACIS), Salvador, Brazil, 15–18 October
2019, pp.788–793. New York: IEEE Computer Society.
The author(s) disclosed receipt of the following financial sup- 12. Lovell DJ. Kinematics-enabled lossless compression of
port for the research, authorship and/or publication of this freeway and arterial vehicle trajectories. J Intell Transport
article: This work was funded by public research projects of S 2019; 23(5): 452–476.
Spanish Ministry of Economy and Competitivity (MINECO), 13. Ribeiro de, Almeida D, de Souza Baptista C, Gomes de,
reference TEC2017-88048-C2-2-R. Andrade F, et al. A survey on big data for trajectory ana-
lytics. ISPRS Int J Geo-Inf 2020; 9(2): 88.
ORCID iDs 14. Wang S, Bao Z, Culpepper JS, et al. A survey on trajec-
tory data management, analytics, and learning, http://
Daniel Amigo https://orcid.org/0000-0001-7138-5508 arxiv.org/abs/2003.11547 (14 December 2020, accessed
David Sánchez Pedroche https://orcid.org/0000-0001- 31 March 2021).
8912-5165 15. Amigo D. Trajectory summarization review analysis,
https://danielamigo.github.io/trajectorySummarisation
References Review/
16. Guo T, Yan Z and Aberer K. An adaptive approach for
1. Zheng Y. Trajectory data mining: an overview. ACM T online segmentation of multi-dimensional mobile data.
Intel Syst Tec 2015; 6(3): 1–41. In: Proceedings of the 11th ACM international workshop
2. European GNSS Supervisory Authority. Power-efficient on data engineering for wireless and mobile access
positioning for the Internet of Things: merging GNSS with (MobiDE’12), Scottsdale, AZ, 20 May 2012, p.7. New
low-power connectivity solutions (white paper). Luxem- York: ACM Press.
bourg: Publications Office of the European Union, 2020. 17. Hansuddhisuntorn K and Horanont T. Improvement of
3. Meratnia N and De By RA. Spatiotemporal compression TD-TR algorithm for simplifying GPS trajectory data.
techniques for moving point objects. In: Bertino E, In: Proceedings of the 2019 1st international conference on
Christodoulakis S, Plexousakis D, et al. (eds) Advances in smart technology and urban development (STUD), Chiang
database technology (EDBT 2004). Berlin; Heidelberg: Mai, Thailand, 13–14 December 2019, pp.1–6. New
Springer, 2004, pp.765–782. York: IEEE Computer Society.
4. Keogh E, Chu S, Hart D, et al. An online algorithm for 18. Leiva LA and Vidal E. Warped K-means: an algorithm
segmenting time series. In: Proceedings of the 2001 IEEE to cluster sequentially-distributed data. Inform Sciences
international conference on data mining, San Jose, CA, 29 2013; 237: 196–210.
November–2 December 2001, pp.289–296. New York: 19. Li L, Xia X, Liu X, et al. Batched trajectory compression
IEEE Computer Society. algorithm based on hierarchical grid coordinates. In:
22 International Journal of Distributed Sensor Networks

Proceedings of the 2019 IEEE 10th international confer- 33. Siddique C and Ban J. Self-adaptive online trajectory
ence on software engineering and service science sampling (SAOTS) using spectral domain properties.
(ICSESS), Beijing, China, 18–20 October 2019, pp.414– Transp Res Proc 2019; 38: 874–893.
418. New York: IEEE Computer Society. 34. Feldman D, Sugaya A and Rus D. An effective coreset
20. Liu J, Li H, Yang Z, et al. Adaptive Douglas–Peucker compression algorithm for large scale sensor networks.
algorithm with automatic thresholding for AIS-based In: Proceedings of the 11th international conference on
vessel trajectory compression. IEEE Access 2019; 7: information processing in sensor networks (IPSN’12),
150677–150692. Beijing, China, 16–20 April 2012. New York: ACM
21. Palma AT, Bogorny V, Kuijpers B, et al. A clustering- Press.
based approach for discovering interesting places in tra- 35. Garcia J, Concha OP, Molina JM, et al. Trajectory clas-
jectories. In: Proceedings of the 2008 ACM symposium on sification based on machine-learning techniques over
applied computing (SAC’08), Fortaleza, Brazil, 16–20 tracking data. In: Proceedings of the 2006 9th interna-
March 2008, p.863. New York: ACM Press. tional conference on information fusion, Florence, 10–13
22. Richter K-F, Schmid F and Laube P. Semantic trajectory July 2006, pp.1–8. New York: IEEE Computer Society.
compression: representing urban movement in a nutshell. 36. Besada J, De Miguel G, Soto A, et al. TRES: multiradar-
J Spat Inf Sci 2012; 4(4): 3–30. multisensor data processing assessment using opportunity
23. Soares A Jr, Moreno BN, Times VC, et al. GRASP-UTS: targets. In: Proceedings of the 2008 IEEE radar confer-
an algorithm for unsupervised trajectory segmentation. ence, Rome, 26–30 May 2008, pp.1–6. New York: IEEE
Int J Geogr Inf Sci 2015; 29(1): 46–68. Computer Society.
24. Soares A Jr, Times VC, Renso C, et al. A semi-supervised 37. Garcia J, Besada Portas JA, Molina JM, et al. Model-
approach for the semantic segmentation of trajectories. based trajectory reconstruction with IMM smoothing
In: Proceedings of the 2018 19th IEEE international con- and segmentation. Inform Fusion 2015; 22: 127–140.
ference on mobile data management (MDM), Aalborg, 38. Kamalzadeh H, Ahmadi A and Mansour S. A shape-
25–28 June 2018, pp.145–154. New York: IEEE Com- based adaptive segmentation of time-series using particle
puter Society. swarm optimization. Inform Syst 2017; 67: 1–18.
25. Song R, Sun W, Zheng B, et al. PRESS: a novel frame- 39. Bellman RE. On the approximation of curves by line seg-
work of trajectory compression in road networks. Proc ments using dynamic programming. Commun ACM 1961;
VLDB Endow 2014; 7(9): 661–672. 4(6): 284–286.
26. Wang T. An online data compression algorithm for tra- 40. Bellman RE and Kotkin B. On the approximation of
jectories (An OLDCAT). Int J Inf Educ Technol 2013; 3: curves by line segments using dynamic programming –
480–487. II, 1962, https://www.rand.org/content/dam/rand/pubs/
27. Yuan G, Zhu M, Qiao S, et al. Sparse high-noise GPS research_memoranda/2008/RM2978.pdf
trajectory data compression and recovery based on 41. Cao W and Li Y. DOTS: an online and near-optimal tra-
compressed sensing. IEICE T Fund Electr 2018; E101-A: jectory simplification algorithm. J Syst Software 2017;
811–821. 126: 34–44.
28. Zhao Y, Shang S, Wang Y, et al. REST: a reference- 42. Imai H and Iri M. Polygonal approximations of a curve
based framework for spatio-temporal trajectory com- – formulations and algorithms. Mach Intell Patt Rec
pression. In: Proceedings of the 24th ACM SIGKDD 1988; 6: 71–86 (also published In: Toussaint GT (ed.)
international conference on knowledge discovery and data Computational morphology, vol. 6. North-Holland
mining, London, 19–23 August 2018, pp.2797–2806. Publishing Company, 1988, pp.71–86.).
New York: ACM Press. 43. Chan WS and Chin F. Approximation of polygonal
29. Katsikouli P, Sarkar R and Gao J. Persistence based curves with minimum number of line segments. In: Ibar-
online signal and trajectory simplification for mobile aki T, Inagaki Y, Iwama K, et al. (eds) Algorithms and
devices. In: Proceedings of the 22nd ACM SIGSPATIAL computation, vol. 650 (ed Goos G and Hartmanis J; Lec-
international conference on advances in geographic infor- ture notes in computer science). Berlin; Heidelberg:
mation systems (SIGSPATIAL’14), Dallas, TX, 4–7 Springer, 1992, pp.378–387.
November 2014, pp.371–380. New York: ACM Press. 44. Chen M, Xu M and Franti P. A fast O(N) multiresolu-
30. Rana R, Hu W, Wark T, et al. An adaptive algorithm for tion polygonal approximation algorithm for GPS trajec-
compressive approximation of trajectory (AACAT) for tory simplification. IEEE T Image Process 2012; 21(5):
delay tolerant networks. In: Marrón PJ and Whitehouse 2770–2785.
K (eds) Wireless sensor networks, vol. 6567 (Lecture notes 45. Daescu O. New results on path approximation. Algorith-
in computer science). Berlin; Heidelberg: Springer, 2011, mica 2004; 38(1): 131–143.
pp.33–48. 46. Daescu O and Mi N. Polygonal chain approximation: a
31. Rana R, Yang M, Wark T, et al. SimpleTrack: adaptive query based approach. Comput Geom 2005; 30(1): 41–58.
trajectory compression with deterministic projection 47. Gotsman R and Kanza Y. A dilution-matching-encoding
matrix for mobile sensor networks. IEEE Sens J 2015; compaction of trajectories over road networks. GeoInfor-
15(1): 365–373. matica 2015; 19(2): 331–364.
32. Siddique C and Ban XJ. State-dependent self-adaptive 48. Kellaris G, Pelekis N and Theodoridis Y. Map-matched
sampling (SAS) method for vehicle trajectory data. Trans- trajectory compression. J Syst Software 2013; 86(6):
port Res C: Emer 2019; 100: 224–237. 1566–1579.
Amigo et al. 23

49. Kolesnikov A. Fast algorithm for ISE-bounded polygo- 63. Cui G, Bian W and Wang X. Hidden Markov map
nal approximation. In: Proceedings of the 2008 15th IEEE matching based on trajectory segmentation with heading
international conference on image processing, San Diego, homogeneity. GeoInformatica 2021; 25: 179–206.
CA, 12–15 October 2008, pp.1013–1016. New York: 64. Ji Y, Zang Y, Luo W, et al. Clockwise compression for
IEEE Computer Society. trajectory data under road network constraints. In: Pro-
50. Lange R, Dürr F and Rothermel K. Online trajectory ceedings of the 2016 IEEE international conference on big
data reduction using connection-preserving dead reckon- data (Big Data), Washington, DC, 5–8 December 2016,
ing. In: Proceedings of the 5th international ICST confer- pp.472–481. New York: IEEE Computer Society.
ence on mobile and ubiquitous systems: computing, 65. Kolesnikov A. Efficient online algorithms for the polygo-
networking and services, Dublin, 21–25 July 2008. Brus- nal approximation of trajectory data. In: Proceedings of
sels: Institute for Computer Sciences, Social-Informatics the 2011 IEEE 12th international conference on mobile
and Telecommunications Engineering (ICST). data management, Lulea, 6–9 June 2011, pp.49–57. New
51. Latecki LJ and Lakämper R. Polygon evolution by ver- York: IEEE Computer Society.
tex deletion. In: Nielsen M, Johansen P, Olsen OF, et al. 66. Lee J-G, Han J and Whang K-Y. Trajectory clustering: a
(eds) Scale-space theories in computer vision, vol. 1682 (ed partition-and-group framework. In: Proceedings of the
G Goos, J Hartmanis and J Van Leeuwen; Lecture notes 2007 ACM SIGMOD international conference on manage-
in computer science). Berlin; Heidelberg: Springer, 1999, ment of data (SIGMOD’07), Beijing, China, 11–14 June
pp.398–409. 2007, vol. 12. New York: ACM Press.
52. Nibali A and He Z. Trajic: an effective compression sys- 67. Lin X, Ma S, Zhang H, et al. One-pass error bounded
tem for trajectory data. IEEE T Knowl Data En 2015; trajectory simplification. Proc VLDB Endow 2017; 10:
27(11): 3138–3151. 841–852.
53. Panagiotakis C, Pelekis N, Kopanakis I, et al. Segmenta- 68. Liu J, Zhao K, Sommer P, et al. Bounded Quadrant Sys-
tion and sampling of moving object trajectories based on tem: error-bounded trajectory compression on the go. In:
representativeness. IEEE T Knowl Data En 2012; 24(7): Proceedings of the 2015 IEEE 31st international
1328–1343. conference on data engineering, Seoul, South Korea, 13–
54. Pulshashi IR, Bae H, Choi H, et al. Simplification and 17 April 2015, pp.987–998. New York: IEEE Computer
detection of outlying trajectories from batch and stream- Society.
ing data recorded in harsh environments. ISPRS Int J 69. Liu J, Zhao K, Sommer P, et al. A novel framework for
Geo-Inf 2019; 8(6): 272. online amnesic trajectory compression in resource-
55. Wu F, Fu K, Wang Y, et al. A graph-based min-# and constrained environments. IEEE T Knowl Data En 2016;
error-optimal trajectory simplification algorithm and its 28: 2827–2841.
extension towards online services. ISPRS Int J Geo-Inf 70. Meng Q, Yu X, Yao C, et al. Improvement of OPW-TR
2017; 6(1): 19. algorithm for compressing GPS trajectory data. J Inf Pro-
56. Zhang Y, Shi G, Li S, et al. Vessel trajectory online cess Syst 2017; 13(3): 533–545.
multi-dimensional simplification algorithm. J Navigation 71. Alvares LO, Bogorny V, Kuijpers B, et al. A model for
2020; 73(2): 342–363. enriching trajectories with semantic geographical infor-
57. Zhao P, Zhao Q, Zhang C, et al. CLEAN: frequent mation. In: Proceedings of the 15th annual ACM interna-
pattern-based trajectory spatial-temporal compression on tional symposium on advances in geographic information
road networks. In: Proceedings of the 2019 20th IEEE systems (GIS’07), Seattle, WA, 7–9 November 2007, p.1.
international conference on mobile data management New York: ACM Press.
(MDM), Hong Kong, China, 10–13 June 2019, pp.605– 72. Pan W, Yao C, Li X, et al. An online compression algo-
610. New York: IEEE Computer Society. rithm for positioning data acquisition. Informatica 2014;
58. Chen C, Ding Y, Wang Z, et al. VTracer: when online 38: 339–346.
vehicle trajectory compression meets mobile edge com- 73. Patroumpas K, Alevizos E, Artikis A, et al. Online event
puting. IEEE Syst J 2020; 14: 1635–1646. recognition from moving vessel trajectories. GeoInforma-
59. Deng Z, Han W, Wang L, et al. An efficient online tica 2017; 21(2): 389–427.
direction-preserving compression approach for trajectory 74. Patroumpas K, Pelekis N and Theodoridis Y. On-the-fly
streaming data. Future Gener Comp Sy 2017; 68: 150–162. mobility event detection over aircraft trajectories. In: Pro-
60. Long C, Wong RC-W and Jagadish HV. Direction-pre- ceedings of the 26th ACM SIGSPATIAL international
serving trajectory simplification. Proc VLDB Endow conference on advances in geographic information systems,
2013; 6(10): 949–960. Seattle, WA, 6–9 November 2018, pp.259–268. New York:
61. Bashir M, Ashraf J, Habib A, et al. An intelligent linear ACM Press.
time trajectory data compression framework for smart 75. Potamias M, Patroumpas K and Sellis T. Sampling tra-
planning of sustainable metropolitan cities. T Emerg Tel- jectory streams with spatiotemporal criteria. In: Proceed-
ecommun T, https://onlinelibrary.wiley.com/doi/abs/ ings of the 18th international conference on scientific and
10.1002/ett.3886 (10 February 2020, accessed 27 April statistical database management (SSDBM’06), Vienna,
2020). 3–5 July 2006, pp.275–284. New York: IEEE Computer
62. Bermingham L and Lee I. A framework of spatio- Society.
temporal trajectory simplification methods. Int J Geogr 76. Resheff YS. Online trajectory segmentation and summary
Inf Sci 2017; 31: 1128–1153. with applications to visualization and retrieval. In:
24 International Journal of Distributed Sensor Networks

Proceedings of the 2016 IEEE international conference on 90. Balzano W and Del Sorbo MR. CoTracks: a new lossy
big data (Big Data), Washington, DC, 5–8 December compression schema for tracking logs data based on mul-
2016, pp.1832–1840. New York: IEEE Computer Society. tiparametric segmentation. In: Proceedings of the 2011 1st
77. Reumann K and Witkam A. Optimizing curve segmenta- international conference on data compression, communica-
tion in computer graphics. In: Proceedings of the interna- tions and processing, Palinuro, 21–24 June 2011, pp.168–
tional computing symposium, 7 January 1974, https:// 171. New York: IEEE Computer Society.
jglobal.jst.go.jp/en/detail?JGLOBAL_ID=201002064588 91. Etemad M, Soares A, Etemad E, et al. SWS: an unsuper-
396801 vised trajectory segmentation algorithm based on change
78. Sánchez-Heres LF. Simplification and event identification detection with interpolation kernels. GeoInformatica
for AIS trajectories: the equivalent passage plan method. 2021; 25: 269–289.
J Navigation 2019; 72(2): 307–320. 92. Etemad M, Soares A Jr, Rose J, et al. A trajectory seg-
79. Sheng K, Liu Z, Zhou D, et al. Research on ship classifi- mentation algorithm based on interpolation-based change
cation based on trajectory features. J Navigation 2018; detection strategies, 2019, http://rgdoi.net/10.13140/
71(1): 100–116. RG.2.2.34157.03049 (accessed 26 June 2020).
80. Tang J, Liu L and Wu J. A trajectory partition method 93. Etemad M, Etemad Z, Soares A, et al. Wise sliding win-
based on combined movement features. Wirel Commun dow segmentation: a classification-aided approach for
Mob Com 2019; 2019: 7803293. trajectory segmentation, http://arxiv.org/abs/2003.10248
81. Trajcevski G, Cao H, Scheuermanny P, et al. On-line data (23 March 2020, accessed 28 June 2020).
reduction and the quality of history in moving objects 94. Gao M and Shi G-Y. Ship spatiotemporal key feature
databases. In: Proceedings of the 5th ACM international point online extraction based on AIS multi-sensor data
workshop on data engineering for wireless and mobile using an improved sliding window algorithm. Sensors
access (MobiDE’06), Chicago, IL, 25 June 2006, p.19. 2019; 19(12): 2706.
New York: ACM Press. 95. Guo Q, Wang H, He J, et al. Graphic simplification and
82. Yan Z, Giatrakos N, Katsikaros V, et al. SeTraStream: intelligent adjustment methods of road networks for navi-
semantic-aware trajectory construction over streaming gation with reduced precision. ISPRS Int J Geo-Inf 2020;
movement data. In: Pfoser D, Tao Y and Mouratidis K, 9(8): 490.
et al. (eds) Advances in spatial and temporal databases, 96. Han Y, Sun W and Zheng B. COMPRESS: a compre-
vol. 6849 (Lecture notes in computer science). Berlin; Hei- hensive framework of trajectory compression in road net-
delberg: Springer, 2011, pp.367–385. works. ACM T Database Syst 2017; 42(2): 1–49.
83. Yin H, Gao H, Wang B, et al. Efficient trajectory com- 97. Kim J. Feature-first add-on for trajectory simplification
pression and queries, http://arxiv.org/abs/2007.04503 (13 in lifelog applications. Sensors 2020; 20(7): 1852.
October 2020, accessed 26 January 2021). 98. Liu X, Lin Z and Wang H. Novel online methods for
84. Zheng Y, Zhang L, Ma Z, et al. Recommending friends time series segmentation. IEEE T Knowl Data En 2008;
and locations based on individual location history. ACM 20(12): 1616–1626.
T Web 2011; 5(1): 1–44. 99. Marino DL and Manic M. Fast trajectory simplification
85. Ke B, Shao J, Zhang Y, et al. An online approach for algorithm for natural user interfaces in Robot program-
direction-based trajectory compression with error bound ming by demonstration. In: Proceedings of the 2016 IEEE
guarantee. In: Li F, Shim K, Zheng K, et al. (eds) Web 25th international symposium on industrial electronics
technologies and applications, vol. 9931 (Lecture notes in (ISIE), Santa Clara, CA, 8–10 June 2016, pp.905–911.
computer science). Cham: Springer, 2016, pp.79–91. New York: IEEE Computer Society.
86. Ke B, Shao J and Zhang D. An efficient online approach 100. Muckell J, Hwang J-H, Patil V, et al. SQUISH: an
for direction-preserving trajectory simplification with online approach for GPS trajectory compression. In:
interval bounds. In: Proceedings of the 2017 18th IEEE Proceedings of the 2nd international conference on com-
international conference on mobile data management puting for geospatial research and applications (COM.
(MDM), Daejeon, South Korea, 29 May–1 June 2017, Geo’11), Washington, DC, 23–25 May 2011, pp.1–8.
pp.50–55. New York: IEEE Computer Society. New York: ACM Press.
87. Long C, Wong RC-W and Jagadish HV. Trajectory sim- 101. Muckell J, Olsen PW, Hwang J-H, et al. Compression
plification: on minimizing the direction-based error. Proc of trajectory data: a comprehensive evaluation and new
VLDB Endow 2014; 8(1): 49–60. approach. GeoInformatica 2014; 18(3): 435–460.
88. Zhao Z and Saalfeld A. Linear-time sleeve-fitting polyline 102. Opheim H. Smoothing a digitized curve by data reduc-
simplification algorithms. In: Proceedings of the Auto- tion methods. In: Encarnacao JL (ed.) Eurographics con-
Carto 13, Seattle, WA, 7–10 April 1997, pp.214–223. ference proceedings. Geneva: The Eurographics
Maryland: American Society for Photogrammetry and Association, 1981, pp.127–135.
Remote Sensing (ASPRS). Bethesda: American Congress 103. Su H, Zheng K, Zeng K, et al. Making sense of trajec-
on Surveying and Mapping . tory data: a partition-and-summarization approach. In:
89. Liu K, Li Y, Dai J, et al. Compressing large scale urban tra- Proceedings of the 2015 IEEE 31st international confer-
jectory data. In: Proceedings of the 4th international workshop ence on data engineering, Seoul, South Korea, 13–17
on cloud data and platforms (CloudDP’14), Amsterdam, 13 April 2015, pp.963–974. New York: IEEE Computer
April 2014, pp.1–6. New York: ACM Press. Society.
Amigo et al. 25

104. Su H, Zheng K, Zeng K, et al. STMaker: a system to 118. Popa IS, Zeitouni K, Oria V, et al. Spatio-temporal
make sense of trajectory data. Proc VLDB Endow 2014; compression of trajectories in road networks. GeoInfor-
7(13): 1701–1704. matica 2015; 19(1): 117–145.
105. Sun S, Chen Y, Piao Z, et al. Vessel AIS trajectory 119. Qian H and Lu Y. Simplifying GPS trajectory data with
online compression based on scan-pick-move algorithm enhanced spatial-temporal constraints. ISPRS Int J
added sliding window. IEEE Access 2020; 8: Geo-Inf 2017; 6(11): 329.
109350–109359. 120. Singh AK, Aggarwal V, Saxena P, et al. Performance
106. Tampakis P, Pelekis N, Doulkeridis C, et al. Scalable analysis of trajectory compression algorithms on marine
distributed subtrajectory clustering. In: Proceedings of surveillance data. In: Proceedings of the 2017 international
the 2019 IEEE international conference on big data (Big conference on advances in computing, communications and
Data), Los Angeles, CA, 9–12 December 2019, pp.950– informatics (ICACCI), Udupi, India, 13–16 September
959. New York: IEEE Computer Society. 2017, pp.1074–1079. New York: IEEE Computer Society.
107. Wei Z, Xie X and Zhang X. AIS trajectory simplifica- 121. Tamilmani R and Stefanakis E. Semantically enriched
tion algorithm considering ship behaviours. Ocean Eng simplification of trajectories. Proc Int Cartogr Assoc
2020; 216: 108086. 2019; 2: 1–8.
108. Li H, Kulik L and Ramamohanarao K. Spatio-temporal 122. Yang M, Yan X, Zhang X, et al. Constrained trajectory
trajectory simplification for inferring travel paths. In: simplification with speed preservation. Cartogr Geogr
Proceedings of the 22nd ACM SIGSPATIAL interna- Inf Sc 2020; 47(2): 110–124.
tional conference on advances in geographic information 123. Yuan D and Wang Y. A multi-UAVs’ trajectory data
systems (SIGSPATIAL’14), Dallas, TX, 4–7 November compression method based on 3D-SPM algorithm. In:
2014, pp.63–72. New York: ACM Press. Proceedings of the 2020 39th Chinese control conference
109. Pikaz A and Dinstein I. An algorithm for polygonal (CCC), Shenyang, China, 27–29 July 2020, pp.6874–
approximation based on iterative point elimination. Pat- 6880. New York: IEEE Computer Society.
tern Recogn Lett 1995; 16(6): 557–563. 124. Zhou Y, Huang M, Jiang F, et al. A visualization-
110. Van Hunnik R. Extensive comparison of trajectory sim- oriented trajectory data compression method. In: Pro-
plification algorithms. Utrecht: Utrecht University, 2017, ceedings of the 2019 IEEE international geoscience and
22 pp. remote sensing symposium (IGARSS’2019), Yokohama,
111. Visvalingam M and Whyatt JD. Line generalisation by Japan, 28 July–2 August 2019, pp.3432–3435. New
repeated elimination of points. Cartogr J 1993; 30: York: IEEE Computer Society.
46–51. 125. Liu G, Iwai M and Sezaki K. An online method for tra-
112. Yan Z, Liu Z and Yuan Q. HESAVE: an approach for jectory simplification under uncertainty of GPS. IPSJ
online heuristic GPS trajectory sampling. In: Skuli- Online Trans 2013; 6: 65–74.
mowski AMJ, Sheng Z, Khemiri-Kallel S, et al. (eds) 126. De Vries GKD and Van Someren M. Machine learning for
Internet of vehicles: technologies and services towards vessel trajectories using compression, alignments and domain
smart city, vol. 11253 (Lecture notes in computer sci- knowledge. Expert Syst Appl 2012; 39(18): 13426–13439.
ence). Cham: Springer, 2018, pp.193–207. 127. Zheng Y, Liu L, Wang L, et al. Learning transportation
113. Birnbaum J, Meng H-C, Hwang J-H, et al. Similarity- mode from raw GPS data for geographic applications
based compression of GPS trajectory data. In: Proceed- on the web. In: Proceeding of the 17th international con-
ings of the 2013 4th international conference on computing ference on world wide web (WWW’08), Beijing, China,
for geospatial research and application, San Jose, CA, 21–25 April 2008, p.247. New York: ACM Press.
22–24 July 2013, pp.92–95. New York: IEEE Computer 128. Chen Y, Jiang K, Zheng Y, et al. Trajectory simplifica-
Society. tion method for location-based social networking ser-
114. Douglas DH and Peucker TK. Algorithms for the reduc- vices. In: Proceedings of the 2009 international workshop
tion of the number of points required to represent a line on location based social networks (LBSN’09), Seattle,
or its caricature. Can Cartogr 1973; 10: 112–122. WA, 3 November 2009, p.33. New York: ACM Press.
115. Feng S, Chen L, Ma M, et al. A turning contour main- 129. Kulik L, Duckham M and Egenhofer M. Ontology-dri-
taining method of trajectory data compression. IOP C ven map generalization. J Visual Lang Comput 2005;
Ser Earth Env 2020; 513: 012058. 16(3): 245–267.
116. Hershberger J and Snoeyink J. Speeding up the Dou- 130. Tobler WR. Numerical map generalization (discussion
glas–Peucker line-simplification algorithm. In: Proceed- paper). Michigan Inter-University Community of Math-
ings of the 5th international symposium on spatial data ematical Geographers, 1966, http://www-personal.umi-
handling, Charleston, S.C., USA., 3-7 August 2000. ch.edu/~copyrght/image/micmg/tobler/a/toblera.pdf
Humanities and Social Sciences Computing Lab, Uni- 131. Tobler WR. An update to ‘Numerical Map Generaliza-
versity of South Carolina. tion’. Cartographica 1989; 26(1): 7–25.
117. Lin C-Y, Hung C-C and Lei P-R. A velocity-preserving 132. Vitter JS. Random sampling with a reservoir. ACM T
trajectory simplification approach. In: Proceedings of the Math Software 1985; 11(1): 37–57.
2016 conference on technologies and applications of artificial 133. Ramer U. An iterative procedure for the polygonal
intelligence (TAAI), Hsinchu, Taiwan, 25–27 November approximation of plane curves. Comput Vision Graph
2016, pp.58–65. New York: IEEE Computer Society. 1972; 1(3): 244–256.
26 International Journal of Distributed Sensor Networks

134. Perez J-C and Vidal E. Optimum polygonal approxima- engineering, Cancun, Mexico, 7–12 April 2008, pp.140–
tion of digitized curves. Pattern Recogn Lett 1994; 15(8): 149. New York: IEEE Computer Society.
743–750. 151. Cudre-Mauroux P, Wu E and Madden S. TrajStore: an
135. Ray BK and Ray KS. A non-parametric sequential adaptive storage system for very large trajectory data
method for polygonal approximation of digital curves. sets. In: Proceedings of the 2010 IEEE 26th international
Pattern Recogn Lett 1994; 15(2): 161–167. conference on data engineering (ICDE’2010), Long
136. Chung K-L, Yan W-M and Chen W-Y. Efficient algo- Beach, CA, 1–6 March 2010, pp.109–120. New York:
rithms for 3-D polygonal approximation based on LISE IEEE Computer Society.
criterion. Pattern Recogn 2002; 35: 2539–2548. 152. Lin C-Y, Chen H-C, Chen Y-Y, et al. Compressing tra-
137. Opheim H. Fast data reduction of a digitized curve. jectories using inter-frame coding, 2010, p.25, https://
Geo-Processing 1982; 2: 33–40. www.iis.sinica.edu.tw/file/entry/8056/FULLTEXT/zh/tr
138. Gao C, Zhao Y, Wu R, et al. Semantic trajectory com- 10007.pdf
pression via multi-resolution synchronization-based 153. Lovell DJ. Lossless compression of all vehicle trajec-
clustering. Knowl-Based Syst 2019; 174: 177–193. tories in a common roadway segment. Comput-Aided
139. Wang Z, Yuan G, Pei H, et al. Unsupervised learning Civ Inf 2018; 33(6): 481–497.
trajectory anomaly detection algorithm based on deep 154. Xu D, Wang Y, Jia L, et al. Compression algorithm of
representation. Int J Distrib Sens N. Epub ahead of print road traffic spatial data based on LZW encoding. J Adv
4 December 2020. DOI: 10.1177/1550147720971504. Transport 2017; 2017: 1–13.
140. Visvalingam M. Cartographic information systems 155. Cao H and Wolfson O. Nonmaterialized motion infor-
research group. Hull: University of Hull, 1992, p.20. mation in transport networks. In: Eiter T and Libkin L
141. Li X, Feng Z, Li Y, et al. Spatio-temporal vessel trajec- (eds) Database theory: ICDT 2005, vol. 3363 (ed Hutchi-
tory smoothing using empirical mode decomposition son D, Kanade T, Kittler J, et al; Lecture notes in com-
and wavelet transform. In: Proceedings of the 2019 puter science). Berlin; Heidelberg: Springer, 2004,
IEEE 4th international conference on big data analytics pp.173–188.
(ICBDA), Suzhou, China, 15–18 March 2019, pp.106– 156. Zheng K, Zhao Y, Lian D, et al. Reference-based frame-
111. New York: IEEE Computer Society. work for spatio-temporal trajectory compression and
142. Yang X, Stewart K, Tang L, et al. A review of GPS tra- query processing. IEEE T Knowl Data En 2020; 32:
jectories classification based on transportation mode. 2227–2240.
Sensors 2018; 18(11): 3741. 157. Li Z, Lee J-G, Li X, et al. Incremental clustering for tra-
143. Feng T and Timmermans HJP. Transportation mode jectories. In: Kitagawa H, Ishikawa Y, Li Q, et al. (eds)
recognition using GPS and accelerometer data. Trans- Database systems for advanced applications. Berlin;
port Res C: Emer 2013; 37: 118–130. Heidelberg: Springer, 2010, pp.32–46.
144. Pelekis N, Tampakis P, Vodas M, et al. In-DBMS 158. Consultative Committee for Space Data Systems
sampling-based sub-trajectory clustering. In: Proceed- (CCSDS). Lossless data compression. (CCSDS 1210-
ings of the 20th international conference on extending B-3). Washington, DC: CCSDS, 2020.
database technology, Venice, 21–24 March 2017. 159. Hatanaka Y. A compression format and tools for GNSS
OpenProceedings.org. observation data. Bull Geogr Surv Inst 2008; 55: 21–30.
145. Kolesnikov A and Fränti P. Reduced-search dynamic 160. Huang Y, Li Y, Zhang Z, et al. GPU-accelerated com-
programming for approximation of polygonal curves. pression and visualization of large-scale vessel trajec-
Pattern Recogn Lett 2003; 24(14): 2243–2254. tories in maritime IoT industries. IEEE Internet Things
146. Keogh E, Chu S, Hart D, et al. Segmenting time series: a 2020; 7(11): 10794–10812.
survey and novel approach. In: Kandel A, Bunke H and 161. Zhao L and Shi G. A method for simplifying ship tra-
Last M (eds) Data mining in time series databases (Series jectory based on improved Douglas–Peucker algorithm.
in machine perception and artificial intelligence). Singa- Ocean Eng 2018; 166: 37–46.
pore: World Scientific Publishing, 2004, pp.1–21. 162. Amigo D, Sánchez Pedroche D, Garcı́a J, et al. Segmen-
147. Shatkay H and Zdonik SB. Approximate queries and tation optimization in trajectory-based ship classifica-
representations for large data sequences. In: Proceedings tion. In: Proceedings of the 15th international conference
of the 12th international conference on data engineering, on soft computing models in industrial and environmental
New Orleans, LA, 26 February–1 March 1996, pp.536– applications (SOCO), Burgos, 16–18 September 2020,
545. New York: IEEE Computer Society. p.10. Cham: Springer.
148. Lange R, Dürr F and Rothermel K. Efficient real-time 163. Zhang S, Liu Z, Cai Y, et al. AIS trajectories simplifica-
trajectory tracking. VLDB J 2011; 20(5): 671–694. tion and threshold determination. J Navigation 2016;
149. Hunter J and McIntosh N. Knowledge-based event 69(4): 729–744.
detection in complex time series data. In: Horn W, Sha- 164. Guerrero JL, Berlanga A, Garcı́a J, et al. Piecewise lin-
har Y, Lindberg G, et al. (eds) Artificial intelligence in ear representation segmentation as a multiobjective opti-
medicine, vol. 1620 (ed Goos G, Hartmanis J and Van mization problem. In: De Leon F, De Carvalho AP,
Leeuwen J; Lecture notes in computer science). Berlin; Rodrı́guez-González S, et al. (eds) Distributed computing
Heidelberg: Springer, 1999, pp.271–280. and artificial intelligence. Berlin; Heidelberg: Springer,
150. Lee J-G, Han J and Li X. Trajectory outlier detection: a 2010, pp.267–274.
partition-and-detect framework. In: Proceedings of the 165. Fikioris G, Patroumpas K and Artikis A. Optimizing vessel
2008 IEEE 24th international conference on data trajectory compression. In: Proceedings of the 2020 21st
Amigo et al. 27

IEEE international conference on mobile data management 168. Tu E, Zhang G, Rachmawati L, et al. Exploiting AIS
(MDM), 2020, pp.281–286, https://www.computer.org/cs data for intelligent maritime navigation: a comprehen-
dl/proceedings-article/mdm/2020/09162228/1m6hFt6gO52 sive survey from data to methodology. IEEE T Intell
166. Shuang S, Yan C and Jinsong Z. Trajectory outlier detec- Transp 2018; 19(5): 1559–1582.
tion algorithm for ship AIS data based on dynamic differ- 169. Sánchez Pedroche D, Amigo D, Garcı́a J, et al. Archi-
ential threshold. J Phys Conf Ser 2020; 1437: 012013. tecture for trajectory-based fishing ship classification
167. Zheng Y, Xie X and Ma W-Y. GeoLife: a collaborative with AIS data. Sensors 2020; 20(13): 3782.
social networking service among user, location and tra-
jectory. IEEE Data Eng Bull 2010; 33: 32–39.