Mutuality-oriented reconstruction and prediction hybrid network for video anomaly detection.

Anomaly detection in surveillance videos aiming to locate anomalous events is a very challenging task, since when training a detector, only normal video samples are available. And thus, most existing approaches address this problem in a semi-supervised way by either predicting or reconstructing the video frames and then compute anomaly scores by comparing the generated frames with reference frames. However, reconstruction-based methods usually lead to mis-detection due to the excessively powerful reconstruction abilities yet the incapable capturing of temporal information, while prediction-based methods are able to perceive temporal information but insufficient to produce realistic future frames. To overcome these problems, we propose a novel Mutuality-Oriented Reconstruction and Prediction Hybrid Network (MORPH-Net) for detecting anomalous events. In the MORPH-Net, a new Mutuality-oriented Training (MO-Training) mechanism is introduced to better combine the advantages of prediction-based models and reconstruction-based models. Compared to traditional single training mechanism or simple fusion mechanism, the MO-Training mechanism can prompt the generator module to produce temporally discriminative and realistic frames which benefit the anomaly detection. The experiments evaluated on three large-scale video surveillance datasets show the efficacy of our method compared with the state-of-the-art approaches. [ABSTRACT FROM AUTHOR]