1) Description:
Computer vision based vehicle detection and tracking, in different light and traffic condition. Comparison against "Ground Truth" generated from laserscanner data.

On the left:
Detection and tracking of the proposed method.
Bounding box contour color is unique for each id; constant contour color means constant id, persistent tracking.
White bounding box contour denotes a detection not yet validated.
Filling color coding:
Green: Detection for which a matching GT' BBox has been found (True positive)
Red: Detection for which a matching GT' BBox has not been found (False positive)
Yellow: Ignored (not considered in the statistics) detection because target not visible from the laserscanner point of view. (See figure 3 from the paper)

On the right:
Laserscanner generated ground truth (GT').
Bounding box contour color is unique for each id.
Diagonal cross (saltire) marks a car, vertical (Greek) cross marks a truck.
GT' bbox are generated, during the matching process, from world coordinates considering a flat ground assumption and fixed extrinsic camera parameter, but compensating the vehicle oscillations (pitching) by using the pitch angle that frame by frame allows the best matching with the computer vision detections.
Filling color coding:
Green: GT' BBox for which a matching detection has been found (True positive)
Red: GT' BBox for which a matching detection has not been found (False negative) 
Yellow: Ignored GT' BBox because not (~fully) visible from the camera

Video:
First 4 seconds show (on the right side of the traffic scene) an example of target visible from the camera but not from the laserscanner (see fig. 3 from the paper); the target is marked in yellow, and ignored in the statistics.
The remainder of the video show results in different traffic situation and varying lighting condition.
Between second 40 and 50 it is possible to observe on the left an example of far away dark vehicle not correctly annotated through laserscanner, which causes the misclassification of a correct detection into a false positive. In the paper, statistics are separated by distance.
Note on classification for statistics:
True positives: visible in both images (green filling)
False positives: visible in left image (red filling)
False negatives: visible in right image (red filling)


2) Multimedia player information:
Format 1280x384, mpeg4 codec, 10fps, 78seconds.
Created with:
mencoder "mf://*.jpg" -mf type=jpg:fps=10 -o output.mpg -ofps 30 -ovc lavc -lavcopts vcodec=mpeg4:vbitrate=690


3) Author contact information
Claudio Caraffi, PhD
Tel: +32 (0)2 712 4567
Fax: +32 (0)2 712 8800
E-mail: <name>.<surname>@toyota-europe.com
        <name>.<surname>@gmail.com
Toyota Motor Europe
Technical Centre
Hoge Wei 33B
B-1930 Zaventem, Belgium