A Closed-loop Background Subtraction Approach for Multiple Models based Multiple Objects Tracking

Normally visual surveillance systems are based on background subtraction to detect foreground objects and then conduct multiple objects tracking with data association and tracking filters in an open-loop procedure. Different from the state-of-the-art approaches, this paper discusses a closed-loop object detection and tracking method. In our proposed method, each pixel is first modeled with an adaptive Gaussian Mixture Models (GMMs). Second, foreground moving objects are tracked by Multiple Hypotheses Trackers (MHT) together with an adaptive Interacting Multiple Models (IMM) method. With the IMM approach, object’s dynamic properties can be better modeled to get more accurate dynamic tracking results. Third, our proposed closedloop approach uses the object tracking results to adjust the GMMs’ parameters to extract foreground object pixels more accurately. With this closed-loop approach, the accuracy of both object detection and tracking are improved without increasing computational costs. The proposed new algorithm is tested with extensive experimental videos collected from different scenarios such as urban streets, intersections, and highways. Experimental results demonstrated the efficiency and robustness of our proposed algorithm in handling object detection and tracking in real-time.