Development of a robust monocular-based vehicle detection and tracking system

This dissertation investigates the techniques for monocular-based vehicle detection. A novel system that can robustly detect and track the movement of vehicles in the video frames is proposed. The system consists of three major modules: a symmetry based object detector for vehicle cueing, a two-class support vector machine (SVM) classifier for vehicle verification and a Kalman filter based vehicle tracker. For the cueing stage, a technique for rapid detection of all possible vehicles in the image is proposed. The technique exploits the fact that most vehicles’ front and rear views are highly symmetrical in the horizontal axis. First, it extracts the symmetric regions and the high symmetry points in the image using a multi-sized symmetry search window. The high symmetry points are then clustered and the mean locations of each cluster are used to hypothesize the locations of potential vehicles. From the research, it was found that a sparse symmetry search along several scan lines on a scaled-down image can significantly reduce the processing time without sacrificing the detection rate. Vehicle verification is needed to eliminate the false detections picked up by the cueing stage. Several verification techniques based on template matching and image classification were investigated. The performance for different combinations of image features and classifiers were also evaluated.