Traffic scenes invariant vehicle detection

Abstract

Although lots of vehicle detection methods can implement vehicle detection with high performance, most of their application is confined by traffic scenes. The detection precision may change heavily with traffic congestion extent, illumination variance and vehicle moving speed. To overcome the problem of weak traffic scene adaptability, a robust vehicle detection method is proposed using the inter-relationship of consecutive multiframes. The changing of frame content is a process including abrupt and gradual variation caused by the objects' color and intensity changing. Thus, the local maxima of consecutive frames' objective function are constructed to determine the best vehicle detection frame. This function is invariant to traffic congestion and vehicle speed, and avoids vehicle segmentation from frames. For illumination invariance, traditional threshold method is substituted by peak searching method. Experiments show that the proposed method implements stably in different traffic scenes than traditional methods, and with the real-time performance and higher detection precision.