A collaborative multi-agent framework for abnormal activity detection in crowded areas

It is common to use Close Circuit Television (CCTV) cameras for the purpose of monitoring in urban areas. Because of laborious nature of the task, the human operators tend to lose attention level, thus causing a possibility of missing important events. The intelligent video surveillance systems can help human operators in performing automatic analysis of video feed for suspicious events. Most of the existing systems require segmenting individuals from the scenes for interpreting their actions. However, segmentation is usually not possible in high density crowded scenes. Moreover, there is a lack of work on automated generation of a collaborative view in a multi-camera network of CCTV cameras. In this paper, we propose an agent based framework for the detection of suspicious activities in crowded scenes in a distributed multi-camera CCTV network environment. The proposed scheme does not require segmentation of individuals from the scene and is thus insensitive to the crowd density. The use of Multi-Agent paradigm has incorporated decentralization, autonomy, fault tolerance and flexibility. We have evaluated our framework on our own generated dataset, a web dataset and a standard dataset from University of Minnesota. The results are promising and show the potential of our framework to work in real environments.