Abstract: Video anomaly detection is one of the most significant research tasks in computer vision area. It aims to intelligently identify the events that do not conform to expected behavior based on pattern recognition and computer vision methods. Video anomaly detection is widely applied and there is an enormous potential demand in modern society. Meanwhile, inspired by the successful achievements in various area of emerging deep learning technologies, more and more newly-emerged methods are conducted on video anomaly detection problem. Firstly, we retrospect the definition and main challenges of vi- deo anomaly detection. Secondly, we introduce the mainstream video anomaly detection methods from three primary technical steps (video event extraction, video event representation, video event modeling and detection) of video anomaly detection, and conclude their advantages as well as drawbacks respectively. Finally, we introduce the benchmark datasets and evaluation metrics of video anomaly detection, compare the performance of mainstream methods and give conclusions and prospects.

Key words: video anomaly detection, machine learning, artificial intelligence, foreground extraction, feature extraction, representation learning, modeling normal events