Abstract
A new algorithm for individual vehicle velocity estimation from a roadside camera sequence is proposed. The authors have developed a new vehicle detection and tracking algorithm using a rectangular model. Detected rectangles generated by background image subtraction technique have one-to-one correspondence with vehicles. To cope with the detection and quantification of errors in velocity estimation, the detected vehicle rear-end bottom positions are converted to those on the road. The velocity can be estimated by the weighted linear least-mean square method from a series of frames. The results show a relationship that estimated longitudinal velocity errors will be proportional to the square of the distance from the camera.