TripVista: Triple Perspective Visual Trajectory Analytics and Its

Application on Microscopic Traffic Data at a Road Intersection

Hanqi Guo 1,2 Zuchao Wang 1 Bowen Yu 1 Huijing Zhao 1 Xiaoru Yuan 1,2∗

1) Key Laboratory of Machine Perception (Ministry of Education), and School of EECS, Peking University, Beijing, P.R. China
2) Center for Computational Science and Engineering, Peking University, Beijing, P.R. China

Figure 1: Interface of Triple Perspective Visual Trajectory Analytics (TripVista) visualizing traffic trajectory data at a road intersection. (a) Spatial
traffic view showing geometrical trajectory information; (b) Temporal views of ThemeRiver and scatterplots; (c) Parallel coordinates plot showing
multiple properties of the multi-dimensional data; (d) Time sliders for two-level time range selection; (e) Control panel for system parameter
settings and data classification.

A BSTRACT to provide access to multiple perspectives for users. Experiments

In this paper, we present an interactive visual analytics system, show that our system is capable of effectively finding both regular
Triple Perspective Visual Trajectory Analytics (TripVista), for ex- and abnormal traffic flow patterns.
ploring and analyzing complex traffic trajectory data. The users Keywords: Visual analytics, Linked view, Spatiotemporal, Multi-
are equipped with a carefully designed interface to inspect data dimensional data visualization, Traffic visualization
interactively from three perspectives (spatial, temporal and multi- Index Terms: I.3.6 [Computer Graphics]: Methodology and
dimensional views). While most previous works, in both visualiza- techniques—Interaction Techniques; H.5.2 [Information Interfaces
tion and transportation research, focused on the macro aspects of and Presentation]: User Interfaces—Graphical User Interfaces
traffic flows, we develop visualization methods to investigate and (GUI)
analyze microscopic traffic patterns and abnormal behaviors. In
the spatial view of our system, traffic trajectories with various pre- 1 I NTRODUCTION
sentation styles are directly interactive with user brushing, together
with convenient pattern exploration and selection through ring-style With rapid economic growth, large increases in both motorization
sliders. Improved ThemeRiver, embedded with glyphs indicating and urbanization have been witnessed in many countries since the
directional information, and multiple scatterplots with time as hori- beginning of the last century. As a consequence, the considerably
zontal axes illustrate temporal information of the traffic flows. Our increased number of automobiles in many cities has quickly con-
system also harnesses the power of parallel coordinates to visual- gested transportation pathways. Heavy traffic may give rise to air
ize the multi-dimensional aspects of the traffic trajectory data. The pollution, aggravate time consumption of people who need to com-
above three view components are linked closely and interactively mute to work or school, pose a significant safety risk, and even ex-
acerbate feelings of inequities in the society. In late August, 2010,
∗ e-mail:{hanqi.guo, zuchao.wang, bowen.yu, zhaohj, the world witnessed the longest traffic jam in China [34]. Road
xiaoru.yuan}@pku.edu.cn construction, traffic accidents and breakdowns created a 60-mile,
11-day bumper-to-bumper gridlock on National Expressway 110
IEEE Pacific Visualisation Symposium 2011 between the capital Beijing and inner Mongolia. It is indisputable
1 - 4 March, Hong Kong, China that transportation has become a key issue in city development and
978-1-61284-934-8/11/$26.00 ©2011 IEEE management in today’s world, which leads to the urgent require-

163
 ment of advanced technologies for us to monitor, model, and opti- moving objects can be detected and tracked to estimate their sta-
mize traffic flows in our transportation systems. tus parameters, including location, speed and direction at each time
Recently, much work has been devoted to improving trans- instance [37]. The advent of GPS and cellphone tracking methods
portation systems with the state-of-the-art information technolo- also makes those properties more accessible [24, 10].
gies. Sensing devices, such as RFID tags, video recorders, laser Trajectory and movement data have been studied with various
scanners, GPS tracking units and cellphones have been utilized to approaches, including visual analysis [3], machine vision [31],
collect various data and construct intelligent transportation systems. clustering [5], feature extraction [4] and movement pattern taxon-
With more and more data available, one major challenge coming omy [14]. Visual analysis tools enable interactive and intuitive data
into our sight is how to effectively analyze the traffic data we have exploration. Andrienko and Andrienko [2] investigate ways of us-
gathered and extract insights. Many researchers have focused on ing aggregation for visual analysis of movement data. A variety
macro simulation data or observations in order to find network bot- of visualizations and interaction techniques are designed to repre-
tlenecks. As a matter of fact, it is also essential to analyze and mon- sent results of aggregations and enable comprehensive exploration
itor the micro behaviors of traffic, find out the origin of accidents of the data. Mosaic diagrams are suggested for the exploration of
and jams, and evaluate traffic light and crossing configurations for cyclical traffic patterns and directional bar diagrams are proposed
potential adjustments. In addition to the traffic scenario that can for the study of movements in different directions. Visualizations
be generated by micro-simulation software like Paramics [12] and play vital roles in finding significant locations, extracting tracks and
VisSim[13], data collected from real traffic flow is vital to research exploring movement dynamics [1]. Anomalies and other activities
as there are many features and exceptions of real-life situations that can also be detected and analyzed by exploiting the intuition and
cannot be well modeled. Micro traffic data is a collection of object experience of security and surveillance experts through an easy-
movements, including the position, speed, size and other proper- to-use visual feedback loop [21]. Viewing datasets from different
ties of the vehicles and pedestrians. The movement of an object is perspectives is of high importance to trajectory data investigation.
usually described as a specific trajectory. Movement data has been studied in geospatial and spatiotemporal
In this work, we analyze the traffic data at a road intersection visualization systems [22, 36]. Those systems provide connections
collected through several laser scanners and some other auxiliary between different perspectives to enhance the capability of visu-
devices, which is a typical microscopic traffic dataset. The dataset alizations. Slingsby [30] proposes a treemap cartography method
is generated from the raw point cloud by post processing tech- to show spatiotemporal traffic patterns. To select a few interest-
niques [37]. Laser-scanned data provides a possibility of study- ing trajectories from a large number, Bouvier and Oates [9] sug-
ing micro behaviors of individual vehicles and pedestrians. Much gest staining and Hurter [19] proposes a brush-pick-drop interac-
more information can be uncovered from such detailed data in com- tion scheme. Their methods are general for 2D trajectory data, but
parison with data obtained through video or other traditional tech- with limited perspectives provided. In our work, we provide com-
nologies, if appropriate analytic tools are provided. However, such prehensive perspectives for data exploration. Our system mainly
datasets are challenging for various reasons. First, at a busy road in- focuses on specific types of data, such as traffic data collected at a
tersection, thousands of moving objects can be captured in a short road intersection.
period of time. It is not a trivial task visualizing and analyzing Several visualization metaphors have been included in our sys-
such datasets efficiently and conveniently. Second, the collection tem. Parallel coordinates [20] have been developed for multi-
of the point clouds inherently includes noise, containing short, in- dimensional visualization. The ThemeRiver metaphor introduced
complete, even nonsense trajectories that can hardly be filtered out by Harve et al. [18] provides an intuitive way for time-varying data
by automatic algorithms. Therefore new approaches are demanded visualization with clustering information. Byron et al. [11] have a
to deal with the magnitude of data and to reduce the influence of detailed discussion on the geometry and aesthetics of ThemeRiver
the inherent noise. as one kind of stacked graph. Wei et al. [33] develop the method
In this paper, we design a visual analytics system, Triple Per- of putting labels into the space midst the ThemeRiver in order to
spective Visual Trajectory Analytics (TripVista), for exploring mi- show text information on email contents. In our work we embed
croscopic traffic data, as illustrated in Figure 1. The developed vi- glyphs into ThemeRiver to illustrate time-varying characteristics
sualization system enables the user to investigate trajectories from as well as directional patterns. Glyph representations have been
different angles, including spatial, temporal and multi-dimensional commonly applied in information visualization [32]. Visualization
perspectives. In our system, by taking advantages of the linked tools can provide interfaces for visually-driven data clustering [28].
views, the user can perceive the underlying features of the dataset Schreck et al. [29] propose a visual-interactive monitoring and con-
and filter out noise and irrelevant trajectories for further investiga- trolling framework extending the basic Kohonen Feature Map al-
tion of interesting cases. Experiments show that our system is ca- gorithm for trajectory clustering. Coordinated views have been
pable of effectively finding regular patterns and anomalies of traffic widely applied to provide effective visualizations and user inter-
flows. actions [8, 16]. In our system, visualizations of spatial, temporal
The remainder of this paper is organized as follows. First the and multi-dimensional perspectives are linked together to provide
related works are reviewed in Section 2. Our design philosophy, to- visual analytics from multiple aspects simultaneously.
gether with the data to be visualized, is given in Section 3. Detailed
description of the proposed visual analytics system is in Section 4, 3 OVERVIEW
followed by several analysis results from the traffic data presented In this section, we first describe the traffic dataset to be explored by
in Section 5. Important issues of the work and possible improve- our proposed visual analytics system. Then the design philosophy
ments are discussed in Section 6 before the paper is concluded in of the proposed system is introduced.
Section 7.
3.1 Data Description
2 R ELATED W ORKS In this work, we focus on a microscopic trajectory dataset collected
Traffic data acquisition, collection and processing have been widely at a road intersection. The data is captured with roadside laser scan-
studied in intelligent traffic system research. Micro traffic data can ners employed to profile the road conditions horizontally from dif-
be obtained by simulation from existing software like VisSim [13] ferent viewpoints [37]. The traffic direction and the traffic light
and Paramics [12]. In recent years, laser scanners and camera- configuration of the cross are shown in Figure 3(a). It is a T-shape
based methods have been developed to capture vehicle data, thus crossroad with a joint one-way road and another multi-lane two-

164
 properties, traffic trajectory data has multi-dimensional attributes
such as object type, path length and direction. Extra information
can also be derived from the raw data. For example, minimum speed
and maximum speed of each individual track can be extracted from
original speed information. We include original and derived dimen-
sions in the parallel coordinates plot for interactive user manipula-
tion. The multi-dimensional perspective also helps to filter out data
noise, and to find interesting patterns.
Linking all the above three perspectives for analysis is a design
Linked of consequence. Any separation would lead to visualization gap
and result in information loss or an incomplete understanding of the
data. Duplex interactions among all views have been established.

4 I NTERFACE
The interface of the system consists of the multiple coordinated
views shown in Figure 1: the traffic view for displaying spatial in-
formation, the ThemeRiver with embedded glyphs and scatterplots
for showing temporal variations, and the parallel coordinates for
Figure 2: Design Philosophy of Triple Perspective Visual Trajectory multi-dimensional visualization. Two-level time sliders supporting
Analytics (TripVista). Spatial, temporal and multi-dimensional per- quick time range selection are integrated to augment data explo-
spectives are closely linked and respectively represented by different ration. In the following, we will introduce design details of each
metaphors. view in our visual analytics system.

4.1 Traffic View

way road. In part of the two-way road, lanes with different di- The major function of traffic view is to directly display spatial infor-
rections are separated with a safety island. The compass indicates mation by rendering each trajectory as a polyline. Each trajectory
the north direction on the data map. The contour points of mov- is drawn according to its scanned position. The traffic view shows
ing objects are captured on a horizontal plane at a scanning rate of the accumulation of all trajectories by blending the tracks as semi-
26ms per frame. When the objects entered the intersection, they transparent polylines (Figure 3(b)). This gives the user an intuitive
were detected, tracked and further classified as cars, buses, bicy- overview of the data over the selected time span. The color of each
cles, pedestrians and others. The dataset contains 209,426 trajecto- line indicates object type: pedestrian, bus, car, bicycle or others.
ries represented by consecutive sampled points recorded over two Alternatively, considering the fact that speed information is also es-
days. There are 33,362,651 sampled points in total and each point sential to pattern study, we provide another shading mode called
has the attributes position, speed, direction and timeInstance. This speed, in which one trajectory is drawn with gradually changing
dataset collected from a real situation is very challenging to be ana- colors. Red color denotes low speed while green denotes higher
lyzed due to its noisy nature and uncertainties caused by occlusions. speed as shown in (Figure 3(c)). A map of the intersection is dis-
The noises resulting from tracking confusions and errors are mainly played as background. Users may recognize general clusters of tra-
in the form of those with an impossibly small faction of movement. jectories through an intuitive perception of the plot. Box-like rep-
We filter out the trajectories that have very short passing distance resentations in the view denote moving objects at a specified time
(<10 meters) or time (<1 second). Through a carefully designed instance (determined interactively in the temporal views or by key-
visual analytics system, we enable direct exploration and manipu- board play/rewind hotkeys).
lation on such a dataset. Convenient selections are integrated into the traffic view. In ad-
dition to normal brushing, the user may apply directional brushing
3.2 Design Philosophy to pick up trajectories with a specific shape by sketching. Further-
Based on the nature of the traffic trajectory data, we design a vi- more, ring-style sliders are overlaid on top of the trajectory paths.
sual analytics system, Triple Perspective Visual Trajectory Analyt- The inner ring slider is used to select the entrance direction, while
ics (TripVista) with three perspectives embedded: spatial, temporal the outer ring is for exiting angle filtering. By adjusting the slider
and multi-dimensional respectively to study microscopic patterns bars, the user can specify patterns with a specific entrance and exit
and discover abnormal behaviors in this data. As illustrated in Fig- range. Histograms showing the density of the trajectories at dif-
ure 2, each perspective has its own visual representation. They are ferent angles are mapped along the ring circumference, which pro-
closely linked to provide powerful visual exploration capabilities. vides additional information concerning the traffic under investiga-
Spatial Perspective Spatial information is essential for under- tion. When the mouse hovers over the histograms, the correspond-
standing object movements in physical space. To identify certain ing region will be shown in an expanded mode to display more
types of object movements in traffic flows, flexible and thorough details. The histogram information of selected trajectories will be
exploration of the geometrical information is indispensable. In our highlighted with a darker color (Figure 3(b)). Rotation of the view
design, spatial perspective (the traffic view) provides the intuitive is also supported for the user’s convenience.
geometrical information.
Temporal Perspective Traffic movements are typically time- 4.2 ThemeRiver Embedded with Glyphs
varying data. Observation of the traffic variation along the temporal Although the standard ThemeRiver can bring forth flow volume in-
axis is vital for identifying corresponding patterns and features, and formation, it alone cannot show details of direction-related traffic
especially for making operational decisions. The temporal perspec- patterns. We design a ThemeRiver view with embedded glyphs in
tive of the visualization is able to provide both an overview span- TripVista to display directional information. Arrow-shape glyphs
ning a long time period and details at a specific time point. We representing specific directional movement of objects are integrated
have employed ThemeRiver with glyphs together with scatterplots into the ThemeRiver. The ThemeRiver view supports convenient
to illustrate temporal features of the dataset. user interactions, e.g., mouse-hovering highlighting , glyph brush-
Multi-dimensional Perspective In addition to the spatiotemporal ing and zooming. These interactions can help the user get intuitive

165
 20m/s
Bicycle
Car
Bus
Pedestrian
Others

0m/s
(a) (b) (c)

Figure 3: Traffic View: (a) The map of the road intersection where
the data is collected. The arrows on the map indicate the permitted
traffic directions and the traffic light configurations; (b) Traffic view -
colored according to object types; (c) Traffic view - colored according (a) (b)
to speed variation.

information both on the flow volume and the flow direction of each
traffic object type. By browsing along the time axis, the user may
easily compare different groups of trajectories and simultaneously
identify patterns through the embedded glyphs. In terms of our de-
sign philosophy, the ThemeRiver with embedded glyphs presents a
way of connecting different perspectives such as direction (spatial),
volume statistics (multi-dimensional) and its time-varying features
(temporal) are coordinately visualized.
In our implementation of the ThemeRiver, two layers are com-
puted: standard ThemeRiver as background, and added glyphs.
Harve’s algorithm [18] is adopted to draw the background The- (c) (d)
meRiver (Figure 4(a)). The glyphs are overlaid on the background
ThemeRiver according to the following criteria: 1) the glyphs Figure 4: Algorithm Illustration for glyph embedding in ThemeRiver:
should clearly and faithfully represent local characteristics of each (a) Original ThemeRiver; (b) Possible glyph positions determined by
traffic type; 2) the glyphs should be distributed evenly; 3) each Fast Hierarchical Importance Sampling; (c) Each river is subdivided
glyph should be placed inside one specific river for aesthetics and and the same type of glyphs are located in the same subriver; (d)
unambiguity; 4) local patterns of glyphs should remain consistent Resulting ThemeRiver with embedded glyphs.
when zoomed in and out. To determine the positions of glyphs, Fast
Hierarchical Importance Sampling [26] with uniform importance is mapping. Scatterplots serve as a medium between the temporal and
applied. As this sampling method has the blue noise property, the the multi-dimensional perspective. The user can brush out a group
sampled points will be distributed randomly and evenly. The points of trajectories quickly based on their occurring time or clustering in
outside the rivers or intersecting the boundaries are discarded. The the scatterplots.
red dots in Figure 4(b) show the computed candidate glyph posi-
4.3 Parallel Coordinates
tions. To decide the type of the glyph at each point, every river is
subdivided into several subrivers according to the directional clus- Parallel coordinates are integrated into TripVista to support multi-
tering of trajectories. Each sampled point is placed exactly inside dimensional data exploration. As one of the most widely applied
one subriver (Figure 4(c)). Thus one type of glyph represents one methods of multi-dimensional visualization, parallel coordinates
directional pattern. Due to the fact that the directional information amplify the system’s ability of viewing data in a multi-dimensional
at a road intersection is naturally described by entrance and exit, for sense. We set the dimensions of parallel coordinates to the en-
simplification we use discrete encoding for the directional patterns trance time, passing time/distance, minimum/maximum/average
and glyph generations rather than a fully-developed clustering algo- speed, start/end bearing, angle change (for both raw and prepro-
rithm based on trajectory comparisons. The details of this process cessed trajectories), maximum/minimum acceleration and object
will be discussed in Section 6.2. After the types of glyphs are de- type. Through parallel coordinates, the user can perceive more fea-
termined, the enhanced ThemeRiver is drawn with corresponding tures of trajectories rather than be limited to basic spatiotemporal
glyphs rendered at each sampled position, as shown in Figure 4(d). properties. The user can clearly see any singularities among the
When zoomed in or out, the sampling frequency of the time points common patterns if the singularities have any distinctive property.
for flux calculation will adapt to the current observation level. The Generally, brushing on the parallel coordinates offers a convenient
changes of river contours in this procedure are usually minor so that way to select those trajectories with interesting features. Noise can
the drawing coherence is maintained. We use the local densities of also be filtered out with parallel coordinates.
glyphs to show the importance of the direction pattern feature in-
stead of size [33]. Each traffic trajectory has a certain group of 4.4 User Interactions
corresponding glyphs, which guarantees the uniqueness of repre- Each view in TripVista supports convenient interactions. Besides
sentation and also enables lasso selections in the ThemeRiver. click and brush, pertinent interactions such as directional brush and
Five scatterplots are also included in the temporal view on ring sliders in the traffic view are provided. A free-form lasso is
the right of the ThemeRiver to enhance TripVista, where each automatically recognized as a rectangle when the moving distance
point represents one trajectory path. The horizontal coordinate of the user’s mouse is close to the straight-line distance between
of each point in the plot is determined by the occurring time of the beginning and ending points of the lasso [25]. Selection re-
the trajectory. Vertical axes are the total passing time, the mini- sults of any of the views support intersection, union and subtraction
mum/maximum/average speed, and the moving distance of the ob- operations. The user may apply group operations to the current se-
jects in each plot respectively. To balance point density and maxi- lection, e.g. they may create or modify groups through the control
mize space utilization, we performed logarithmically scaled distor- panel, which provides a way for user-defined classification. User-
tion transformation on the vertical axes [27] in addition to the linear defined color schemes for groups are supported for multi-purpose

166
suitability. The two time sliders can be used to select a target time with spatial understanding through brushing. Then the user can go
range at lower and higher levels, to which flow volume histograms into more detailed information, such as speed, object type and oth-
are mapped for a quick overview of distributions. In detail, the ers through operation with linked views. We emphasize the internal
lower-positioned time slider performs on the higher level — en- connections between space and other perspectives, as well as the
abling the selection of time spans in tens of minutes or hours from natural perceptional progressing from one to another.
the traffic on July 16th, 2008 (Figure 1), while the upper-positioned
time slider functions at a lower level — enabling the selection of
time ranges in seconds and minutes. The ThemeRiver view may be
switched to scatterplots by moving the thumbnails of scatterplots
in a drag-and-drop manner in order to assign the scatterplots more
space for display when necessary. We leave the decisions on the
parameters including rendering transparency of trajectories and ob-
jects, size of brushed and unbrushed points in the scatterplots, scale
of histograms, etc. to the users.
(a) (b) (c)
5 V ISUAL A NALYSIS R ESULTS
By using TripVista, not only can the user obtain overview informa- Figure 5: Investigation of U-turn Patterns (a) Select the U-turn trajec-
tion of traffic flow over a long time period, but they may also drill tories with ring sliders; (b) Select the U-turn trajectories using direc-
down to traffic trajectory details and discover interesting micro be- tional brush; (c) Replay the scene (U-turn patterns in red).
haviors or patterns through interactions. Three cases are demon-
strated in this section to show how TripVista helps analyze traffic 5.2 Case 2: Find Patterns and Violations
flow data from integrated spatial, temporal and multi-dimensional The second case focuses on regular pattern recognition and viola-
perspectives. Our results illustrate the importance of linked per- tion detection.
spectives, which is the key part of our design philosophy. Through TripVista, the user can perceive traffic flow patterns in
multiple angles. The regular traffic light patterns can be discovered
5.1 Case 1: Investigate Specific Behaviors through the temporal view, which can be illustrated by showing
The first case is an example of recognizing special spatial patterns time variation of the vehicle traffic volume. Such an intermittent
in the traffic view with the advantage of ring sliders and the direc- pattern can hardly be observed without the help of the temporal
tional brush. view.
TripVista is capable of identifying different types of traffic tra- We first select the trajectories entering the intersection from the
jectory patterns according to differences in their geometry. Dense lower-left road. The system then splits and regroups the paths into
rendering line bundles in the traffic view provide an intuitive visual two groups according to their exit directions, which can be done
cue of the directions of the trajectory groups. with the ring sliders or the directional brush as we have described
In the dataset we are working on, drivers from the low-left main previously. The result is shown in Figure 6.
street are allowed to make a U-turn at the interaction, as illustrated It is clear in the glyph-embedded ThemeRiver that the left-turn
in Figure 3(a). We are interested in investigating how frequently trajectories (brown river with left-to-top arrows) and the straight
such U-turn actions occurred. With TripVista, we first filter out trajectories (blue river with left-to-right arrows) have a very clear
pedestrians and bicycles by brushing the dimension of type in the temporal patterns. The left-turn traffic light turns green following
parallel coordinates accordingly. The U-turn traffic then can be im- the go-forward traffic light turning into green with a short lag period
mediately isolated by adjusting the ring slider bars to the corre- for the pedestrians to cross the road. Here embedded glyphs help
sponding positions and ranges as shown in Figure 5(a). Alterna- the user quickly identify the flow direction within a certain river
tively, directional brush sketched by the user can also disclose a (traffic type). The difference is derived from the underlying traffic
similar pattern as illustrated in Figure 5(b). light regulation. It can also be perceived in the ThemeRiver view
The selected trajectories are also highlighted accordingly in that the left-turn traffic flow has smaller volume. Through the rivers
other views simultaneously, which enables the user to investigate the proportion variation of the two flows in a single day can be
other properties at the same time in addition to the geometrical in- further investigated.
formation. For example, the user can explore the scatterplots or
the parallel coordinates in other view windows for information of
vehicle speed. The visualization shows that the average speeds of
these trajectory patterns are generally low and the minimum speed
values are nearly zero. There are only four cases in which the min-
imum speed of the vehicle is above 2 m/s. This indicates that most
cars drive cautiously and make a stop during the execution of a U-
turn. Our system also provides play-back function of the selected
scenes. Figure 5(c) shows individual moving cars with box-like
representations after the time span is narrowed to a very short pe-
riod. The system also provides information on the selected traffic
volume. Among the total number of 32,777 moving objects during
the selected time span, cars and buses together account for 9,421, as (a) (b)
illustrated in the top-left corners of the figures. The number of tra-
jectories selected by ring sliders is only 85. This indicates that there Figure 6: Regular Traffic Patterns (a) Group the concerned trajec-
were only 85 cars that made a U-turn in that time period. The direc- tories according to their exits; (b) Selected trajectory groups have
tional brush gives the smaller number of 42 due to more restricted different volumes as shown in the ThemeRiver.
shape matching.
This example demonstrates how to use TripVista to discover and In addition to understanding general patterns, TripVista also en-
investigate specific trajectory patterns. The process usually starts ables us to gain insights into the micro behaviors of pedestrians and

167
 vehicles. Since the vertical road is one-way only as shown in Fig- we select the trajectories with larger values of angle change with the
ure 3(a), any automobile retrograding violates the traffic rule. parallel coordinates interface (Figure 8(a), (b)) and identify one tra-
After brushing these trajectories with the reversed direction of jectory of a bicycle passing the center of the intersection as shown
the one-way road and excluding some obvious noise, the trajecto- in Figure 8(c). By replaying the scene in the traffic view, it can be
ries of the offenders can be shown in Figure 7(a). The cyan trajec- observed that the large angle change actually resulted from the cy-
tories representing cars indicate the violation events. Bicycle traffic clist’s evasive actions to avoid collision with a car. The car made
flows are represented as trajectories in purple. The cyclist’s behav- a panic brake in front of the cyclist in that incident. It can be con-
ior is legal since the applicable regulation in this case only covers cluded that it is a dangerous event (Figure 8(d)). Besides the single
cars. TripVista can also expose another type of offence on this one- violation, it is interesting to reveal through this example that vio-
way road. When brushing the traffic from the right to the left, we lations usually appear in groups. Since the violating cyclist passed
can observe that a few vehicles turned right first at the intersection, directly along the diagonal of the intersection, the travel distance is
and then made an immediate U-turn. At this road intersection, right rather long. We brush similar behaviors by selecting long passing
turns are always allowed regardless of the status of the traffic lights. distances in one of the scatterplots in a short time span covering
Those identified trajectories indicate drivers who wanted in effect the event identified above. Then we find that there were several
to make a left at the intersection during the red light time. which other bicycles going in a similar fashion at the time of the incident
is illegal. Furthermore, with the replaying function in the temporal (Figure 8(e)). Such observation helps us identify social behaviors
view, we can watch the box-like representation executing the entire as violations tend to appear in group due to people’s false sense of
maneuver. The identified violation is shown in Figure 7(b). safety.
This example illustrates the process of identifying both regular
patterns and violations, which is one of the essential objectives of
the TripVista design. Visualization methods in our design present
information from different perspectives, thereby offering more op-
portunities for the user to perform in-depth investigation.
(a)

(a) (b)
(b) (c)
Figure 7: Traffic Violations (a) Wrong-way offenders; (b) Illegal turn-
ing pattern.

5.3 Case 3: Discover Hidden Information

In the third case, all views are used in conjunction to figure out
interesting but usually hidden information like a chain of inci-
dents among the densely mapped trajectories. This exemplifies that
TripVista can be used to detect potentially dangerous cases, and
even to help make conclusions concerning social behaviors.
In traffic flows, there often exist chain reactions in which the
behavior of one object (vehicle or pedestrian) is the result of the ab- (d) (e)
normal behavior of other moving objects. For example, an abrupt
stop of a moving vehicle would result in several following vehicles Figure 8: Dangerous Events Discovered in Dense Trajectories: (a)
slamming on the brakes. However, such chain reactions are usually Selection of the trajectories with larger angle change values in the
hidden among the data collected and cannot be directly identified. parallel coordinates; (b) Trajectories with larger angle changes in the
To discover such events, one needs to fully utilize all the perspec- traffic view; (c) One interesting trajectory identified in the traffic view;
tives including space, time and multi-dimension to view the data (d) A dangerous event: the cyclist narrowly escaped being hit by a
from multiple angles. TripVista offers the possibility of such com- car; (e) Other violation behaviors in groups.
plex investigation. The below example shows an actual finding of a
dangerous moment that was initiated by a cyclist’s violation behav-
ior.
6 D ISCUSSION
We calculate the angle change for each trajectory, which is the
sum of the included angles (always positive) of consecutive speed In this section, we summarize the a few aspects of the implemen-
vectors. Trajectories with smooth behaviors would have angle tation details of TripVista. Important issues including automatic
changes no more than 90 degrees except those performing a U- algorithm features and possible improvements to the system are dis-
turn. An abnormal trajectory that exhibits sudden turns or evasion cussed. User feedback is provided to further confirm the application
actions will have angle changes larger than 90 degrees. In this case, value of the system to its potential users.

168
6.1 Implementation Details and Scalability ing strategy is appropriate and sufficient in the current case due to
The system was developed in C++ with Qt 4.7, boost 1.43.0 and the fact that patterns, as well as abnormal behaviors at a road in-
OpenGL 3.1. The program has been tested on a Dell T3500 work- tersection have limited variation on possible movement directions.
station, with an Intel Xeon W3503 2.4GHz CPU, 2GB RAM and In the future, we would like to extend our approach to more com-
a NVIDIA GeForce GTX 275 graphics card with 896MB memory. plex traffic movements within larger road networks. More com-
User interactions can be performed smoothly at a frame rate of more plex phenomena should be considered, for example the question
than 20 frames/second if middle size data is load. of how a congestion propagates through a network can potentially
To mitigate the high demand on computing resources when vi- be visually analyzed by our system. Some research has been done
sualizing large datasets, intermediate results of all views are saved to cluster general movement, such as probabilistic trajectory clus-
in frame buffers and only necessary parts are updated when small tering [17] and density-based trajectory clustering [15, 7] . Such
changes are made by interactions. For example, the highlighting automatic algorithms can be enhanced by visualization and interac-
effects can be drawn on top of the existing buffer contents. Our tion [28, 29, 5]. It is also possible to include comprehensive cluster-
system can run on regular workstations or PCs with general hard- ing algorithms in our system. Since it is very difficult to predict all
ware settings. possible trajectory patterns, glyphs in our ThemeRiver can be au-
To take advantage of a better user experience during interactions, tomatically generated as a visual summary of the trajectory shapes
we have also tested our system with a Wacom Cintiq 12WX display they correspondingly represent.
tablet as the input device. Interactions such as brushing and hov- Views and interactions are adaptable to be applied in more com-
ering can be performed more efficiently on the tablet with a pen plex conditions. For example, ring sliders can be improved to deal
through sketching than through a mouse and keyboard. The advan- with road networks instead of one single road intersection. We
tages of interactive features of TripVista can be further utilized on are considering enhancing the current system by introducing ring
such devices to enhance data exploration. slices, which are partial ring sliders with more freedom in shapes.
Our system can also be extended to handle 3D scenarios, such as
6.2 Extension of Applicability airline traffic data. The directional brush and ring sliders can be
TripVista helps the user discover interesting patterns by includ- improved to support 3D selections.
ing not only visualization representation techniques, but also semi-
6.3 User Feedback
automatic algorithms such as the directional brush and angle change
calculation. We have consulted people who have expert knowledge in intelligent
More automatic algorithms can be integrated into our system to transportation system research for comments after internally testing
enhance its capability. For example, relative motion detection is an our TripVista system.
interesting topic in microscopic traffic analysis. We have discov- The feedback is very encouraging. The highly supported interac-
ered a case in which a vehicle narrowly avoided colliding with a tions are greatly appreciated by the experts. Especially the powerful
bicycle in Section 5.3. It would be beneficial to automatically ex- selection function as well as the animation replaying the scenes re-
tract similar cases in the whole dataset. One possible solution is to ceive very positive comments. These features largely enhance the
search for matches of a predefined patterns [23], such as encounter user experiences on traffic data exploration. According to the feed-
or approximation [6]. Another is to search for behaviors similar to back, our system is also good for demonstration. The visualization
a given case. We have integrated into our system a prototype of provides intuition for understanding the data.
a relative motion detection algorithm by searching the similar be- Our visual analytics cases are also convincing to the domain ex-
haviors, taking both spatial and temporal information together with perts. The analysis of the traffic flow patterns in Section 5.2 is
neighboring movement information into account. As shown in Fig- considered very valuable to guide the optimization of traffic lights
ure 9, our system can detect a few similar patterns in which moving control. The discovery of the cyclist’s case in Section 5.3 greatly
bicycles and cars appeared very close. The user can further inspect arouses the experts’ attention. The detected event is later further
the detected events and evaluate the degree of danger accordingly. confirmed by their manual inspection on the initial laser-scanned
The detection is also helpful for data correction. As shown in the point cloud data. The experts also indicate that the tablet demon-
last case of Figure 9, one bicycle is inside the space of a bus at a stration of our system shows a data exploration process much more
specific moment. This impossible scene results from inaccuracies efficient and convenient than traditional methods, which they would
of the data. Such findings are extremely valuable for data cleaning like to apply in their work.
and validation.
7 C ONCLUSION AND F UTURE W ORK
In this paper, we present a novel visual analytics system named
TripVista for visualizing microscopic traffic trajectory data at a
road intersection. Based on the triple perspective design philos-
ophy, coordinated multiple visualizations with convenient interac-
tions are created in our proposed system. By simultaneously in-
specting the time-varying trajectory data from spatial, temporal and
multi-dimensional views our system provides powerful tools for
data comprehension and exploration. The establishment of close
Figure 9: Several cases discovered by our prototype of relative mo- linking among each view is essential to the visualization effective-
tion detection algorithm, in which moving bicycles (purple) and cars ness in our system.
(cyan) appeared very close. However, when speed is considered, As illustrated in the examples of traffic data analysis with
none of them is as dangerous as the case in Section 5.3. It is notice- TripVista, our system helps the user not only comprehend regu-
able that in the last case, the bicycle is inside the car. This impossible lar traffic patterns, but also discover abnormal behaviors. Some
scene results from the inaccuracies in the data. possible improvements are named in our user feedback. More au-
tomatic algorithms are expected, such as advanced clustering and
In terms of clustering techniques, our present system clusters the relative motion detections. In the future, we would like to explore
trajectories based on the entrance and exit information. It uses pre- the potential of TripVista by equipping it with improved automatic
defined glyphs to represent different patterns. This simple cluster- algorithms and additionally extending it to more complex data, such

169
 as traffic in a larger road network, road-free movement and 3D tra- [17] S. Gaffney and P. Smyth. Trajectory clustering with mixtures of re-
jectory data. We also plan to integrate more powerful visualization gression models. In KDD’99, pages 63–72, 1999.
techniques. For example, by introducing point representation in the [18] S. Havre, E. G. Hetzler, P. Whitney, and L. T. Nowell. ThemeRiver:
parallel coordinates plots [35] in our system, we can enhance the Visualizing thematic changes in large document collections. IEEE
ability to analyze high dimensional aspects of the trajectory data. Trans. Vis. Comput. Graph., 8(1):9–20, 2002.
[19] C. Hurter, B. Tissoires, and S. Conversy. Fromdady: Spreading air-
craft trajectories across views to support iterative queries. IEEE Trans.
ACKNOWLEDGEMENTS Vis. Comput. Graph., 15(6):1017–1024, 2009.
The authors thank the anonymous reviewers for their invaluable [20] A. Inselberg. The plane with parallel coordinates. The Visual Com-
comments and suggestions. This work is supported by National puter, 1(2):69–91, 1985.
Natural Science Foundation of China Project No.60903062, Bei- [21] F. Janoos, S. Singh, O. Irfanoglu, R. Machiraju, and R. Parent. Activ-
jing Natural Science Foundation Project No.4092021, 973 Program ity analysis using spatio-temporal trajectory volumes in surveillance
Project No.2009CB320903, 863 Program Project 2010AA012400, applications. In Proceedings of IEEE Symposium on Visual Analytics
Chinese Ministry of Education Key Project No.109001 and IIPL- Science and Technology, pages 3–10, 2007.
[22] T. Kapler and W. Wright. GeoTime information visualization. In Pro-
09-016. The work of data collection was supported by NSFC
ceedings of the IEEE Symposium on Information Visualization, pages
Project No.60975061. The authors would also like to thank Jie Sha
25–32, 2004.
for precious comments and feedback. [23] P. Laube, S. Imfeld, and R. Weibel. Discovering relative motion pat-
terns in groups of moving point objects. International Journal of Ge-
R EFERENCES ographical Information Science, 19(6):639–668, 2005.
[1] G. Andrienko and N. Andrienko. Exploration of massive movement [24] G. Leduc. Road traffic data: Collection methods and applications.
data: a visual analytics approach. In AGILE ’08, 2008. Technical report, Institute for Prospective Technological Studies, EU,
[2] G. Andrienko and N. Andrienko. Spatio-temporal aggregation for vi- 2008.
sual analysis of movements. In Proceedings of IEEE Symposium on [25] M. J. McGuffin and I. Jurisica. Interaction techniques for selecting
Visual Analytics Science and Technology, pages 51–58, 2008. and manipulating subgraphs in network visualizations. IEEE Trans.
[3] G. Andrienko and N. Andrienko. A visual analytics approach to ex- Vis. Comput. Graph., 15(6):937–944, 2009.
ploration of large amounts of movement data. In VISUAL ’08: Pro- [26] V. Ostromoukhov, C. Donohue, and P.-M. Jodoin. Fast hierarchical
ceedings of the 10th international conference on Visual Information importance sampling with blue noise properties. ACM Transactions
Systems, pages 1–4, 2008. on Graphics, 23(3):488–495, 2004.
[4] G. Andrienko, N. Andrienko, J. Dykes, S. I. Fabrikant, and M. Wa- [27] S. Potts and T. Möller. Transfer functions on a logarithmic scale for
chowicz. Geovisualization of dynamics, movement and change: key volume rendering. In Proceedings of Graphics Interface 2004, pages
issues and developing approaches in visualization research. Informa- 57–64, 2004.
tion Visualization, 7(3):173–180, 2008. [28] S. Rinzivillo, D. Pedreschi, M. Nanni, F. Giannotti, N. Andrienko, and
[5] G. Andrienko, N. Andrienko, S. Rinzivillo, M. Nanni, D. Pedreschi, G. Andrienko. Visually driven analysis of movement data by progres-
and F. Giannotti. Interactive visual clustering of large collections of sive clustering. Information Visualization, 7(3):225–239, 2008.
trajectories. In Proceedings of IEEE Symposium on Visual Analytics [29] T. Schreck, J. Bernard, T. Tekusova, and J. Kohlhammer. Visual clus-
Science and Technology, pages 3–10, 2009. ter analysis of trajectory data with interactive kohonen maps. In Pro-
[6] N. Andrienko, G. Andrienko, M. Wachowicz, and D. Orellana. Un- ceedings of IEEE Symposium on Visual Analytics Science and Tech-
covering interactions between moving objects. In Extended Abstracts nology, pages 3–10, 2008.
of GIScience 2008, pages 16–26, 2008. [30] A. Slingsby, J. Wood, and J. Dykes. Treemap cartography for showing
[7] M. Ankerst, M. M. Breunig, H.-P. Kriegel, and J. Sander. Optics: spatial and temporal traffic patterns. Journal of Maps, v2010:135–146,
Ordering points to identify the clustering structure. In SIGMOD’99, 2010.
pages 49–60, 1999. [31] J. Vijverberg, N. de Koning, J. Han, P. de With, and D. Cornelissen.
[8] A. Barsky, T. Munzner, J. Gardy, and R. Kincaid. Cerebral: Visu- High-level traffic-violation detection for embedded traffic analysis. In
alizing multiple experimental conditions on a graph with biological ICASSP’07, volume 2, pages 793–796, 2007.
context. IEEE Trans. Vis. Comput. Graph., 14(6):1253–1260, 2008. [32] M. O. Ward. A taxonomy of glyph placement strategies for multidi-
[9] D. J. Bouvier and B. Oates. Evacuation traces mini challenge award: mensional data visualization. Information Visualization, 1(3/4):194–
Innovative trace visualization staining for information discovery. In 210, 2002.
Proceedings of IEEE Symposium on Visual Analytics Science and [33] F. Wei, S. Liu, Y. Song, S. Pan, M. X. Zhou, W. Qian, L. Shi, L. Tan,
Technology, pages 219–220, 2008. and Q. Zhang. Tiara: a visual exploratory text analytic system. In
[10] Y.-J. Byon, B. Abdulhai, and A. Shalaby. Real-time transportation KDD’10, pages 153–162, 2010.
mode detection via tracking global positioning system mobile devices. [34] M. Wines. Chinas 60-mile traffic jam is breaking up,
Journal of Intelligent Transportation Systems, 13:161–170, 2009. 2010. http://wheels.blogs.nytimes.com/2010/08/24/chinas-60-mile-
[11] L. Byron and M. Wattenberg. Stacked graphs - geometry & aesthetics. traffic-jam-is-breaking-up/.
IEEE Trans. Vis. Comput. Graph., 14(6):1245–1252, 2008. [35] X. Yuan, P. Guo, H. Xiao, H. Zhou, and H. Qu. Scattering points in
[12] G. Cameron, B. J. N. Wylie, and D. McArthur. Paramics—moving parallel coordinates. IEEE Trans. Vis. Comput. Graph., 15(6):1001–
vehicles on the connection machine. In Supercomputing’94, pages 1008, 2009.
291–300, 1994. [36] H.-S. Zhang, Y. Zhang, Z.-H. Li, and D.-C. Hu. Spatial-temporal
[13] H. Chen, X. Zhang, and G. Liu. Simulation and visualization of em- traffic data analysis based on global data management using mas.
pirical traffic models using vissim. In Proceedings of IEEE Interna- IEEE Transactions on Intelligent Transportation Systems, 5(4):267–
tional Conference on Networking, Sensing and Control 2007, pages 275, 2004.
879–882, 2007. [37] H. Zhao, J. Cui, H. Zha, K. Katabira, X. Shao, and R. Shibasaki. Sens-
[14] S. Dodge, R. Weibel, and A.-K. Lautenschütz. Towards a taxonomy of ing an intersection using a network of laser scanners and video cam-
movement patterns. Information Visualization, 7(3):240–252, 2008. eras. IEEE Intelligent Transportation Systems Magazine, 1(2):31–37,
[15] M. Ester, H.-P. Kriegel, J. Sander, and X. Xu. A density-based algo- 2009.
rithm for discovering clusters in large spatial databases with noise. In
KDD’96, pages 226–231, 1996.
[16] W. Freiler, K. Matkovic, and H. Hauser. Interactive visual analysis of
set-typed data. IEEE Trans. Vis. Comput. Graph., 14(6):1340–1347,
2008.

170