Bayesian Cellular Automata Fusion Model Based on Dual-Stream Strategy for Video Anomaly Action Detection
Zhongtang Zhao, Ruixian Li · Pattern Recognition and Image Analysis · 2021
Abstract In recent years, public safety issue is paid more attention, and the video anomaly action in the crowd seriously affects the safety of people’s life and property. Therefore, video anomaly detection has become a hotspot in image processing, machine vision, and other related fields. Traditional video anomaly detection methods do not consider the problem of video time information; therefore, this paper proposes a Bayesian cellular automata fusion model based on dual-stream strategy for video anomaly action detection. The spatial flow model uses convolutional self-encoding network to reconstruct a single frame of video. The time flow model uses convolutional long-short-term memory encode-decode network to reconstruct the short-term optical flow sequence. Then, the reconstruction errors of each frame in the space flow model and the time flow model are calculated, respectively. The adaptive threshold is designed to binarize the reconstruction error graph. Based on Bayesian cellular automata fusion criterion, the reconstruction errors of space flow and time flow are fused to obtain the fused reconstruction error graph. On this basis, the anomaly action detection result is obtained. Experimental results on UCSD and Avenue video datasets show that the proposed algorithm is superior to the existing video anomaly detection algorithms.