A Boost Strategy to the Generative Error Based Video Anomaly Detection Algorithms
share
The generation error (GE) based algorithms show excellent performances in the task of video anomaly detection. However, in the step of anomaly detection, they have two problems: (1) Abnormal events usually occur in local areas. It reduces the saliencies of the abnormal events to use the frame-level GE as the anomaly-score. (2) Every discrimination has both advantages and disadvantages, it is difficult to aggregate multiple discriminations effectively. To address these problems, we propose a promotion strategy which is consisted of two modules. Firstly, we replace the frame-level GE with the maximum of the block-level GEs in a frame as the anomaly score. Secondly, assuming that the stricter the anomaly threshold the more reliable the anomaly detected, we propose a reliable-anomaly (R-anomaly) based multiple discriminations aggregation method. In this method, we set a strict anomaly detection threshold (SADT) for each auxiliary discrimination to detect R-anomalies. Then we use the detected R-anomalies to enhance their anomaly scores in the main discrimination. Experiments are carried out on UCSD and CUHK Avenue datasets. The results demonstrate the effectiveness of the proposed strategy and achieve state-of-the-art performance.
READ FULL TEXT
