In this paper, we develop a robust vision-based
approach for real-time traffic data collection at nighttime.
The proposed algorithm detects and tracks vehicles through
detection and location of vehicle headlights. First, we extract
headlights candidates by an adaptive image segmentation
algorithm. Then we group headlights candidates that belong
to the same vehicle by spatial clustering and generate vehicle
hypotheses by rule-based reasoning. The potential vehicles
are then tracked over frames by region search and pattern
analysis methods. The spatial and temporal continuity
extracted from tracking process is used to confirm vehicle's
presence. To handle problem of occlusions, we apply Kalman
Filter to motion estimation. We test the algorithm on the
video clips of nighttime traffic under different conditions.
The experimental results show that real-time vehicle
counting and tacking for multi-lanes are achieved and the
total detection rate is above 96%.