Abnormal event detection based on cosparse reconstruction

Abstract

A novel video abnormal event detection method based on cosparse reconstruction with local self-similarity constraint is proposed. For a given spatio-temporal patch which is represented by a feature vector using concatenated multi-scale histogram of optical flow, abnormal event detection is implemented by cosparse reconstruction with respect to an analysis dictionary learned from normal event set. To adapt to the diversity of normal events, feature space is partitioned into meaningful subspaces by clustering and cosparse sub-dictionary is learned from each cluster. Experimental results show that the proposed approach achieves competitive performance with the state-of-the-art methods.