Collaborative Foreground Background Object Isolation and Tracking

In this chapter, we propose a collaborative framework for efficient isolation and tracking of foreground objects. The algorithm is able to operate in very complex and dynamic background environments. In case that a large deviation of the foreground and/or the background object is encountered, new model classifiers are retrieved or dynamically created to satisfy the current visual characteristics. For this reason, on-line learning classification schemes are incorporated with the purpose of dynamically adjust the performance of a model classifier to the current visual statistics. Object evaluation is accomplished using spatial and temporal criteria. In particular, in case that the motion compensated mask deviates a lot from the current detected mask, the new model selection module is activated. Approximation of the foreground and the background object is performed using the mutual exclusion properties between the two masks as well as the motion information in case that neither the background nor the foreground object is accurate. Experimental results are presented, which indicates the robust foreground detection even in case of complex background content with high dynamic changes.