Using adaptive background subtraction into a multi-level model for traffic surveillance

Interest in the application of computer vision techniques to automatic video-based analysis of traffic is high at present. This is due in part to the capabilities of video sensors, as well as to social demands for traffic safety. In general, these systems are cheaper and less disruptive than other kinds of devices like loop detectors for traffic monitoring. Automatic traffic surveillance is, however, still a challenging problem when we consider many of the practical difficulties involved (i.e. limited number of cameras and positions of these with respect to the scene, variable illumination and weather conditions, intrinsic complexity of analyzed traffic events, need for a real-time frame rate processing, among others). In this paper, we propose a multi-level framework for automatic analysis of complex traffic videos which present different kind of variations. The accurate and efficient extraction of relevant scene information from the video frames is performed in a hierarchical bottom-up form using the system presented. First of all, foreground moving pixels are detected in each frame using a proposed method of adaptive background subtraction. After that, these pixels are grouped into blobs if they share some common properties. Blobs detected in predefined scene entry regions are identified as vehicles and these are tracked along the controlled road area. At the upper level, some traffic monitoring statistics and also related linguistic reports on the evolution of traffic in the scene are generated periodically. Experimental results on the adaptive background method proposed, as well as regarding its integration in the multi-level traffic analysis system, are very satisfactory for the traffic videos analyzed.