Research on Abnormal Behavior Recognition on Campus Based on Dual-stream 3D FasterNet

In recent years, public safety problems frequently occur, especially school violence has become the focus of social attention. Slowfast is a dual-stream difference speed video detection network with a 75% accuracy rate in video motion recognition, but Slowfast has a long running time and insufficient accuracy for the application of school violence detection. In order to solve the problems of slow operation speed and insufficient accuracy of the Slowfast model, this paper proposes to replace the backbone network with the latest lightweight feature extraction network FasterNet. At the same time, in order to ensure the network’s ability to extract time information, the two-dimensional network FasterNet is designed in three dimensions and improved to 3D FasterNet to replace the backbone network 3D Resnet in Slowfast. The utilization efficiency of hardware is improved by 30%, and the overall operation time of the network is reduced by 20%. After the improvement, the average accuracy of the network is improved by 4% and the accuracy of fighting action recognition is improved by 5%.