Vehicle detection and speed estimation using cascade classifier and sub-pixel stereo matching

Abstract

A new vision-based approach to accomplish automatic detection and speed measuring of vehicles is proposed in this paper. In the proposed method, a cascade classifier, based on Haar features, is trained on frontal view of vehicles and deployed for vehicle detection. A fast and accurate foreground segmentation algorithm is proposed to distinguish moving vehicles from the background and prune detection results. The refined detection result in each individual frame is aggregated with tracking results based on a combination of Kalman filter and Munkres assignment algorithm. To accomplish an accurate and real-time speed measurement, an efficient sub-pixel matching technique in conjunction with stereo-vision framework is applied to calculate per-frame displacements of vehicles. Experimental results show that the proposed method leads to more reliable detections and more accurate speed estimations as compared to competing algorithms.