Dual-Branch Framework with Convolutional Attentive Block for Video Anomaly Detection
Qun Li, Rui Yang, Yaying Shen, Ziyi Zhang, Xianzhong Long · 2023
Video anomaly detection is a challenging task due to the diversity and low frequency of anomalous events. Inspired by two-stream paradigms, we propose a hybrid dual-branch framework for improving the robustness to different anomalous scenarios. We integrate the prediction branch and the reconstruction branch, where the prediction branch learns appearance features from video frames, and the reconstruction branch learns motion features from optical flows, which allows the dual-branch framework to be jointly optimized. Specially, a channel attention block is developed for our dual-branch framework to adaptively enhance feature representations by modeling the interdependence among feature channels. Experimental results on three benchmarks prove the effectiveness of our dual-branch framework for video anomaly detection.