Research on abnormal behavior recognition algorithm based on 3D-resnet deep neural network
Q. Ye, Xiao Long Lu · IET conference proceedings. · 2023
Traditional video surveillance requires a large amount of manpower and material resources, requiring manual attendance and only providing a single monitoring function. Under long-term and high traffic data conditions, duty personnel are prone to fatigue and overlook abnormal behaviors of security risks in videos. Therefore, video systems that provide intelligent recognition of abnormal behaviors in videos are particularly important, as they can help security management departments achieve intelligent security management, the abnormal behavior recognition algorithm is the core soul of the system. This article studies the use of deep learning algorithm ``3D residual deep neural network'' (referred to as 3D-Resnet) to identify three types of abnormal behavior risks in video surveillance, including: fighting, robbery, and arson hazards. The designed algorithm is applied in an intelligent recognition system for video abnormal behavior, and the experimental results show that the recognition accuracy is over 90%. The designed algorithm can effectively detect new input videos in the intelligent recognition system for video abnormal behavior, and has a good recognition effect on abnormal behavior recognition. According to experimental comparison, the neural network algorithm of 3D combined residual structure is significantly superior to C3D neural network and Two-stream neural network in identifying abnormal behavior.