Application of Sparse Representation and Camera Collaboration in Visual Surveillance Systems

With the growth of demand for security and safety, video-based surveillance systems have been employed in a large number of rural and urban areas. The problem of such systems lies in the detection of patterns of behaviors in a dataset that do not conform to normal behaviors. Recently, for behavior classification and abnormal behavior detection, the sparse representation approach is used. In this paper, feature sparse representation in a multi-view network is used for the purpose of behavior classification and abnormal behavior detection. To serve this purpose, a geometrically independent feature is first extracted for each location in the image. Then, for each camera view, the matrix for the dictionary A is calculated, which is considered as a set of behavior models. In order to share information and make use of the trained models, the learned dictionary matrix from the experienced camera is transferred to inexperienced cameras. The transferred matrix in the new camera is subsequently used to detect abnormal behaviors. A hierarchical method on the basis of spectral clustering is proposed for learning the dictionary matrix. After sparse feature representation, a measurement criterion, which makes use of the representation, is presented for abnormal behavior detection. The merit of the method proposed in this paper is that the method does not require correspondence across cameras. The direct use of the dictionary matrix and transfer of the learned dictionary matrix from the experienced camera to inexperienced ones, are tested on several real-world video datasets. In both cases, desirable improvements in abnormal behavior detection are obtained. The experimental results point to the efficacy of the proposed method for camera cooperation in order to detect abnormal behaviors.