A constrained sparse representation approach for video anomaly detection

Sparse representation based anomaly detection algorithms have received a widely interest in recent years. However, most of the existing approaches fail to pay attention to the manifold structure of the video data, which has been pointed to be important for data representation. To overcome this limitation, we develop a new sparse coding algorithm named constrained sparse representation (CSR) for video anomaly detection, which explicitly takes the manifold structure of data into account. CSR assumes that each sample's sparse coding coefficient can be linearly reconstructed by the coding coefficients of its neighbors. Therefore, CSR algorithm can obtain relatively smoothly sparse representations along the manifold of data. We apply the proposed CSR model to video abnormal event detection task and conduct extensive experiments on the UMN database. The experimental results demonstrate that our proposed CSR model performed better than some related algorithms.