Improving performance via post track analysis

In this paper, we improve the effective performance of a surveillance system via post track analysis. Our system performs object detection via background subtraction followed by appearance based tracking. The primary outputs of the system however, are customized alarms, which depend on the user´s domain and needs. The ultimate performance therefore depends most critically on the receiver operating characteristic curve of these alarms. We show that strategically designing post tracking and alarm conditions can improve the effective performance of the system improved dramatically. This addresses the most significant error sources, namely, errors due to shadows, ghosting, temporally or spatially missing fragments and many of the false positives due to extreme lighting variations, specular reflections or irrelevant motion.