Vehicle detection and recognition are very important issues in vision-based traffic applications. We are interested in detecting and recognizing moving vehicles in real road scenes captured by a static camera. An image difference based method is adopted for the segmentation of moving vehicles from input road images. Therefore, the exact background image is needed for successful vehicle detection. We used a Kalman filter based concept to obtain and continuously update the background image. We propose a new technique which adaptively determines the Kalman gain by using the Mahalanobis distance. Motion detection process consists of the initial mode for obtaining an approximate background image and the detection mode for segmenting moving vehicles as a next step.
Then recognition is performed only within regions where a moving objects are detected. As an off-line process, we build model set, including three typical vehicle types sedan, jeep, and bus, directly from the input image. In next step, we find the vehicle roof by maximizing an energy function defined by edge, line reliability and corner reliability. Using four lines of the roof as the bases of canonical coordinate we calculate five-line invariant of other lines. Finally we recognize the moving vehicle from the model set using five lines invariants.