Unsupervised Video Anomaly Detection in Traffic and Crowded Scenes
Satoshi Hashimoto, Alessandro Moro, Kenichi Kudo, Takayuki Takahashi, Kazunori Umeda · 2022 IEEE/SICE International Symposium on System Integration (SII) · 2022
In this paper, we propose a scene-independent robust unsupervised video anomaly detection method based on future frame prediction as a breakthrough and better video anomaly detection technique. Most conventional methods evaluate and develop a static camera and dashcam approach as independent tasks, and no method has been proposed that is independent of the capture conditions. The proposed method introduces a frame-wide future prediction-based spatio-temporal adversarial networks that can handle arbitrary series lengths to cope with various scenes. The noise in the prediction error caused by constant background changes is improved by weighting the regions of interest for the discriminator of the generative adversarial networks (GANs). This framework can be applied to all cases regardless of the scene environment. Experiments on public datasets of general traffic scenes and crowded scenes confirm the superiority of the proposed method over current state-of-the-art methods.