0% found this document useful (0 votes)
94 views4 pages

Are To

This document describes a video-based traffic monitoring system that can track multiple vehicles in real-time and classify them without using specialized hardware. The system detects moving objects in a video stream using background subtraction and then segments, tracks, and classifies the objects as vehicles move through the scene. The algorithms were implemented on a PC with a DSP board and can process video at 8 Hz with resolutions up to 256x256 pixels in real-time. Testing showed the system could successfully track vehicles and estimate their speed and position within 3% error for low to medium traffic conditions.

Uploaded by

Basil
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
94 views4 pages

Are To

This document describes a video-based traffic monitoring system that can track multiple vehicles in real-time and classify them without using specialized hardware. The system detects moving objects in a video stream using background subtraction and then segments, tracks, and classifies the objects as vehicles move through the scene. The algorithms were implemented on a PC with a DSP board and can process video at 8 Hz with resolutions up to 256x256 pixels in real-time. Testing showed the system could successfully track vehicles and estimate their speed and position within 3% error for low to medium traffic conditions.

Uploaded by

Basil
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 4

89

VIDEO-BASED TRAFFIC MONITORING


M Kilger

Siemens AG, FRG

ABSTRACT
A video-based system for traffic monitoring (e.g. on highways
or near crossroads) is presented. The objective of the system is
to set up a high-level description of the traffic scene,
comprising the position, speed and class of vehicles in a range
of view between 0 and about 100 meters. The algorithms run
in real-time on a low-cost system (386-PC with a DSP plug-in
board).
INTRODUCTION
Todays road traffic demands more efficient and intelligent
management and control strategies. As a prerequisite, online traffic data acquisition is necessary. Typical
requirements a r e [81:
- wide range ofview of a t least lOOm should be monitored,
- vehicles should be counted, tracked and classified,
- the speed and position of each vehicle should be reported
rather exactly.
Conventional techniques (e.g. inductive loop) cannot provide
these data, if costs have to be kept within reasonable limits.
In this paper a video nsed traffic monitoring system IS
presented. Several other authors ([41,[51,~61.[71.[81in Europe
and some others in Japan) have presented similar work, but
have not fully met the above requirements. No system is able
to track and classify multiple vehicles in real-time with high
resolution and without using special dedicated hardware.
This work shows that real-time traffic monitoring is possible
even with general-purpose hardware.
A series of problems has to be addressed:
- great variability of the possible scenes,
- partial occlusion of vehicles in heavy traffic conditions,
- difficult illumination conditions,
- real-time processing,
- cost aspects.
On the other hand, the envisioned application allows to take
advantage of a couple of assumptions:
- only objects, which are moving or have been moving in the
past have to be registered,
- the movement of the vehicles can be modelled using
simple physical laws,
- the shape of the vehicles can be modelled using simple
geometric models.
- only part of the image (the traffic region) is of interest.

timing requirements and are tailored to different situations.


According to the situation we select the most appropriate of
the algorithms.
ALGORITHMS

Q
image sequence

I detection of moving objects I

cluster resolution

25
ri
object list

tracking

track list
classification

Fig.1: Flow chart


Detection of moving objects
Moving objects are detected very reliably by comparing the
current image with a background image. A robust
background extraction routine is used ([11,[21,[31). The whole
detection algorithm proceeds in several steps:

Information accumulated over a sequence of frames may be


used.
In Fig.1 the main processing steps a r e shown. For some steps
we provide several algorithms, which diiTer in memory and

Subtraction ofcurrent image from background image.


A threshold decision which separates regions with fast
motion from regions with no or slow motion, is made and
delivers a binary object mask. Of course the properties of
the object mask depend critically on the chosen
threshold value. If the threshold is too high, the mask
will often be fragmented (Fig.3). In the opposite case, lots
of uninteresting movements will be detected, like the
rustle of the leaves. This problem can be overcome by
continously adapting the threshold.
Segmentation ofthe object mask.

90

A segment list is built up. For each connected component


the bounding box, the area in the image and an
estimation of the corresponding area in the real world
a r e computed. The latter is derived using the inverse
perspective mapping under the assumption that vehicles
move on the road at height zero.
Removal of segments that are unlikely to be related to
vehicles.
We remove segments which are off the traffic regions
and segments, which are too small.
Update of the background image.
Outside a mask, which corresponds to moving objects in
the image, an adaptation routine is applied. It is based
on recursive filtering with variable coefficients. The
mask consists of the bounding boxes of individual
vehicles (see below). They represent the objects
occluding the background more precisely than the object
mask originally created in step 1.

Fig.3: Object mask

Fig.2: Current image


Fig.4Object mask plus the bounding box
and then matched with a geometric model. As geometric
models for vehicles in the image, bounding boxes turn out to
be sufficient.
Cluster resolution
A special problem is the resolution of clusters into single
vehicles in heavy traffic situations where vehicles may
partially occlude each other. Under the following
assumptions a solution is found:
- daylight conditions, diffuse illumination,
- the camera is positioned to the side of the road, e.g. on a
pole for the traffic light or on a bridge.
The first step is the detection ofclusters within the segments.
If the bounding box is greater than a defined limit, the
segment probably is a cluster.
Next the clusters have to be resolved into single vehicles.
Under our assumptions parts of the leading edges can be seen
as prominent feature ofthe vehicles. These parts are detected

Furthermore the knowledge about the past is utilized. The


vehicles moving off the camera usually have been seen
isolated in the past and only later join into clusters. Hence,
this information can be used. On the other hand. vehicles
which move towards the camera are first seen within clusters,
which then gradually resolve. In this case, the resolution
algorithm, which takes its main information from one image
frame, is effectively used. Plausibility values for the result of
the resolution are built up.
Segments which belong to the same object have to be joined.
Here the same information as for resolving clusters is used.

91

current position,
predicted position for the next image frame,
- current speed,
- predicted speed for the nest image frame,
- width,
- plausibility.
The tracking algorithm comprises the following subtasks:
- Assign detected moving objects to existing tracks:
-

Fig.5: Updated background image

Fig.6: model arrangement


The results ofthis processing step are stored in an object list.

'ig.8: Object mask plus the segments aRer the resolution

Tracking
The main task of the tracking algorithm is to track the
detected objects and to predict their respective positions in
the next image frame. In the track list, the whole description
of the scene is stored. Each track has the following attributes:

The position of an object is defined as the centre of the


leading edge (see above!). The object that is closest to the
predicted position of a track is assigned to t h a t track.
Create new tracks:
A new track is built, if the detected object cannot be
assigned to a n existing track.
Remove tracks:
If, in the current cycle, an object cannot be assigned to an
existing track, the track is extrapolated via a Kalmanfilter prediction. Ifthe extrapolated position is outside the
considered scene, the track is removed. Otherwise the
plausibility of the track is reduced. If the plausibility is
lower than a defined limit, the track is also removed.
Predict the position and the speed for the next cycle:
The position and speed of the tracks are filtered in world
coordinates using a Kalman-filter. The use of Kalmanfilters implies a smoothing of the measurements and
provides a n estimation of position and speed. Experiments
have shown that a constant Kalman gain is sufficient.

Classification
The vehicles have to be classified according to:
- trucks,
- cars,
- motorcyles. bicycles.
The selection of the features best suited for the classification
depends on the situation. If the vehicle is moving towards the
camera, the first feature to evaluate is the width of the
bounding box. This width is tracked over the sequence and an
estimation over several frames is made.
If a rotation of the vehicle is detected by the tracking
algorithm, the leading edge is used as suitable feature. This
leading edge again is tracked over several frames and an
estimation of the width is made. This predicted position is
used for feature tracking.
Another useful feature to evaluate is the height of the
bounding box. These measurements are filtered with a
Kalman-filter and the length and height of the vehicle are
calculated.

IMPLEMENTATION
The following input format turns out to be appropriate:
- resolution: 256 x 256 pixel
- grey levels: 8 bit
- frame rate: 8 1i3 Hz
The algorithms were first implemented and tested on a SUN
Sparcstation 1.
Real-time processing is now done on a personal computer (PC)
with an Intel i386 processor a s host, a digital signal processor
(Motorola DSP96002) plug-in board and a frame grabber. The

92

low-level image processing (background extraction and


segmentation) is done by the DSP. The host operates on the
segment list and carries out the high-level algorithms. In
addition it supervises the communication between the
subsystems.
RESULTS AND CONCLUSION
The algorithms were tested for various traffic and
illumination conditions. For low to medium traffic the
clusters can be successfully resolved. The errors in estimating
position and speed were smaller than 3%
Current work follows several lines:
- Automatic adaptation of parameters (e.g. camera
parameters),
- Learning of the scene geometry,
- Feedback from high-level to low and medium level
routines (e.g. background adaptation and segmentation),
- Switching between various algorithms depending on
illumination conditions.
REFERENCES
111 W.Feiten. A.v.Brandt, G.Lawitzky, LLeuthausser . Ein
zur
Erfassung
von
videobasiertes
System
Verkehrsdaten. Proc. 13th DAGM 1991, Munich, pp.
507-514
121 K.P.Karmann. A.v.Brandt. Moving object segmentation
based on adaptive reference images. Proc. of EUSIPCO
1990, Barcelona
[31 R.Ger1. Detektion bewegter Objekte i n digitalen
Bildfolgen natiirlicher Szenen mit unbewegtem
Hintergrund. 1990, Diplomarbeit TU Munich
I41 A. Bielik. T. Abramczuk. Real-time wide-area traffic
monitoring: information reduction and model-based
approach. Proc. 6th Scandinavian Conference on Image
Analysis, 1989, Oulu, Finland, pp.1223-1230
[51 S.Beucher, J.M.Blosseville, F.Lenoir. Traffic Spatial
Measurements Using Video Image Processing. SPIE
Vo1.848 Intelligent Robots and Computer Vision: Sixth
in a Series (1987). pp. 648-655
[61 N.Hoose, L.G.Willumsen. Automatically extracting
traffic data from videotape using the CLIP4 parallel
image processor. Patt. Recog. Letters, 6 . 1987. pp. 199213
(71 A.D.Houghton, GSHobson, N.L.Seed. R.C.Tozer.
Automatic vehicle recognition. 1989, University of
Sheffield, U.K.
[81 W.Leutzbach,
H.P.Bahr,
U.Becker,
T.Vogtle.
Entwicklung eines Systems zur Erfassung von
Verkehrsdaten
mittels
photogrammetrischer
Aufnahmeverfahren und Verfahren zur automatischen
Bildauswertung. 1987, Technischer Schldbericht,
University Karlsruhe

You might also like