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

Abstract

Normally visual surveillance systems are basedon background subtraction to detect foreground objects andthen conduct multiple objects tracking with data associationand tracking filters in an open-loop procedure. Differentfrom the state-of-the-art approaches, this paper discusses aclosed-loop object detection and tracking method. In ourproposed method, each pixel is first modeled with an adaptiveGaussian Mixture Models (GMMs). Second, foregroundmoving objects are tracked by Multiple Hypotheses Trackers(MHT) together with an adaptive Interacting MultipleModels (IMM) method. With the IMM approach, object’sdynamic properties can be better modeled to get more accuratedynamic tracking results. Third, our proposed closedloopapproach uses the object tracking results to adjust theGMMs’ parameters to extract foreground object pixels moreaccurately. With this closed-loop approach, the accuracy ofboth object detection and tracking are improved withoutincreasing computational costs. The proposed new algorithmis tested with extensive experimental videos collected fromdifferent scenarios such as urban streets, intersections, andhighways. Experimental results demonstrated the efficiencyand robustness of our proposed algorithm in handling objectdetection and tracking in real-time.