Accurate Speed Measurement from Vehicle Trajectories using AdaBoost Detection and Robust Template Tracking

Image streams from road surveillance cameras often show a poor quality for measuring the vehicle´s speed. This paper discusses methods for detection and tracking of vehicles that 1) provides statistical data on bidirectional traffic and 2) provides continuous speed measurements of individual vehicles for a typical range up to 100 m for rear viewed and 70 m for oncoming vehicles, respectively. Detection is based on a block-based variant of the AdaBoost detector with edge orientation histograms. The tracker uses a robust variant of the extended Lucas Kanade template matching algorithm. Results on day and night time sequences show detection on from a vehicle size of 20 x 20 pixels, and tracking down to 10 x 10 pixels, with a speed accuracy of 2.3%, for 95% of the vehicles.