Detecting Abnormal Events Using Locality-Sparse Tensor Reconstruction on Covariance Descriptors

Abnormal event detection in videos is to find spatially or temporally isolated feature points that do not conform to a pre-defined notion of normal ones. Because the covariance descriptors can naturally fuse appearance and motion statistics, resulting in a much smaller dimensionality and compact representation than vector-based features. In this paper, we employ it as event representation, to capture the anomalies that different from normal patterns in appearance and motion simultaneously. Motivated by the application of tensor sparse coding, we propose a novel Locality-Sparse Tensor Reconstruction Model (LTR) to learn the covariance feature distribution for video events. The locality-sparse scheme makes the sample reconstructed only from its neighbourhoods, which bears much less computational complexity and obtain more robustness. To efficiently adapt new video data streams, we employ an online updates for the model. The proposed approach is evaluated on several publicly available datasets and outperforms several methods based on vector feature representation proposed before.