Fig 1.
Flow chart of multi-vehicle detection and tracking.
Fig 2.
Top row: basic forms of Haar-like features. Bottom row: vehicle rear appearances suitable for Haar-like features.
Fig 3.
Cascade Adaboost classifier for vehicle detection.
Fig 4.
Training samples for classifier.
Fig 5.
(a)Sliding window sampling. (b)Raw detection results. (c)Confidence map of raw detections. (d)Final results after redundant boxes fusion.
Fig 6.
(Left) Vanishing point based method fails to filter the false hypotheses with bottom edge under vanish point; (Right) Proposed method successfully filters out the false hypothesis.
Fig 7.
(Left) The visualization of the matching confidence matrix. (Right) Association results between the measurements (yellow) and existing tracks (red).
Fig 8.
The generic track state transitions in the track management.
Fig 9.
Detection results in different conditions.
(a) Dark illumination. (b) Bright illumination. (c) Large scale due to ultra-near distance. (d) Small scale due to long distance. (e) Mottled shadow. (f-g) Slight partial occlusion. (h) Dense traffic. (i)Truck. (j) Van. (k) Strong pose change. (l)Clutter background.
Table 1.
Vehicle detection performance.
Fig 10.
Video sequence of single vehicle with large scale change.
Blue-KF, green-GT, red-ours.
Fig 11.
Tracking performance comparison between KF and AKF.
(a)Tracking error of location. (b)Tracking error of width.
Table 2.
Tracking error comparison between AKF and KF.
Fig 12.
Typical tracking samples of the video captured on 3-Ring road.
(Yellow) Original cascade Adaboost detector. (Red) Ours. Note the detection failure and false alarm of original detector in Frame 61 and 738 respectively.
Fig 13.
The trajectories of different vehicles in 3-Ring road.
(a)The tracking result of our method, and (b)the result of original Cascade Adaboost detector. Note that our method could track each vehicle with correct ID, and efficiently suppress the false positive detections and temporal loss of detection.
Fig 14.
Typical tracking results of the video captured on 5-Ring road.
(Yellow) Original cascade Adaboost detector. (Red) Ours. Note that different vehicles are tracked with independent ID numbers throughout the video by our method. Note the detection failures and false alarms of original detector.
Fig 15.
The trajectories of different vehicles.
(a)The trajectories of different vehicles produced by our method. (b)The detection result generated by original Cascade Adaboost detector. Our method could track each vehicle with correct ID, and efficiently suppress the false positive detections and temporal loss of detection.
Table 3.
Multi-vehicle tracking performance.
Fig 16.
(a)The vehicle in the right lane is not detected due to partial appearance. (b) The vehicle in the ego lane is not detected due to high contrast background.