A video anomaly detection method based on percentile loss training and attention mechanism

Currently, the mainstream methods for video anomaly detection include frame reconstruction and video stream prediction. These methods are always confronted with the problem of over-generalization, which makes the network reconstruct or predict both normal and anomalous modes well and makes it difficult to distinguish the normal and anomalous modes. In this paper, we propose a video anomaly detection model based on the U-Net architecture, in which percentile loss (PL) training is introduced to weaken the generalization of the network by diminishing the influence of complex regular modes that tend to be anomalous on parameter update, and in turn weaken the model’s ability to predict anomalous modes. Moreover, based on the fact that most of the objects in video surveillance are stationary and abnormal behaviors tend to occur in moving objects, a variance attention module is used after the encoder module to make the network focus on moving objects that tend to contain abnormal modes. The proposed framework has been evaluated on publicly available real-world anomaly detection datasets including CUHK Avenue, UCSD and ShanghaiTech. The experiments show that the proposed algorithm improves the detection performance compared with the state-of-the-art algorithms.