Video-Based Traffic Analysis System Using a Hierarchical Feature Point Grouping Approach

This study presents a video-based car detection and tracking algorithm for traffic analysis system. By setting a detection window at the entry of a traffic lane, the algorithm applies a two-layer SVM classifier to detect car-passing events at the detection window. When car-passing event occurs, Harris corners are detected at the detection window and then tracked by optical flow. These tracked corners are then organized by using a hierarchical feature point grouping approach which not only groups the corners which should belong to a single vehicle but also rejects outlier corners. Each vehicle is thus detected and tracked. A variety of traffic parameters such as vehicle count, traffic flow density, vehicle speed detection, and lane change event detection can be further obtained. Experimental results reveal that the proposed method shows better performance than conventional background-subtraction based or virtual-wires based methods on challenging videos such as with traffic jam or at night.