Improving motion state change object detection by using block background context

Abstract

Motion state change object detection, such as stopped objects detection, is one of important topics in Video Surveillance Systems. Generally, backgrounds in the most Video Surveillance Systems have the property of pureness and self-similarity. In this paper, we propose a block background context based background model to solve the motion state change problem. Unlike the classical background model, our approach first models blocks of background, and then determines the learning rate of each block background model by using the block background context information. There are two main advantages. First, the model adaptively selects the learning rate for each block of background model, and that is more flexible than the adaptive learning rate for the whole background. Second, context information helps the determination of true foreground and brings in more reliable information in foreground detection. Our experiments results show that our model outperforms the higher and lower learning rate Gaussian mixture background model in motion state change object detection.