Anomaly Detection in Crowded Scenes via SA-MHOF and Sparse Combination

In this paper, we propose a novel approach to detect video abnormal event which makes a combination of motion and appearance features. For motion feature, we propose a new descriptor called Spatial Associated MHOF (SA-MHOF). SA-MHOF can not only characterize the motion velocity and direction from the optical flow map but also make the detection more accurate. Object far away from the camera suffer from a problem that the magnitude of optical is less than its real speed. We use SA-MHOF based on a spatial related threshold of MHOF to solve this problem. For appearance descriptor, we adopt spatial-temporal gradient to represent the texture feature. To reduce computation cost and achieve fast detection, we adopt sparse combination learning method to model normal events and detect anomaly. We employ a late fusion strategy to make combination of motion and appearance features. Our experiments on the UMN and UCSD datasets show that our approach achieves comparable performance to the state-of-the-art anomaly detection methods.