Multimodal Generative Neural Network for Anomaly Events Detection and Localization in Videos
Mingchen Yang, Shahram Shirani · 2021
Anomaly detection has gained more and more attention with the popularity of automatic surveillance services. However, the variety and uncertainty of abnormal objects and events make the detection and localization of them difficult. In this paper, we propose a Multimodal Generative Neural Network (MGNN), unsupervised anomaly events detection and anomaly objects localization method based on Generative Adversarial Network. Our architecture contains two components, an appearance generation network and a motion generation network. In training, only normal frames are fed into networks. At testing time, since the trained model has learned reconstruction of normal frames, frames with abnormal object will be reconstructed poorly. We use this poor reconstruction to detect abnormal frames. Our experiments with UCSD pedestrian 2 dataset show that our approach achieves 96.5% Area Under Curve (AUC) in frame-level detection and 94.1% AUC in pixel-level detection.