Video anomaly detection based on attention and efficient spatio-temporal feature extraction

An anomaly is a pattern, behavior, or event that does not frequently happen in an environment. Video anomaly detection has always been a challenging task. Home security, public area monitoring, and quality control in production lines are only a few applications of video anomaly detection. The spatio-temporal nature of the videos, the lack of an exact definition for anomalies, and the inefficiencies of feature extraction for videos are examples of the challenges that researchers face in video anomaly detection. To find a solution to these challenges, we propose a method that uses parallel deep structures to extract informative features from the videos. The method consists of different units including an attention unit, frame sampling units, spatial and temporal feature extractors, and thresholding. Using these units, we propose a video anomaly detection that aggregates the results of four parallel structures. Aggregating the results brings generality and flexibility to the algorithm. The proposed method achieves satisfying results for four popular video anomaly detection benchmarks.