Video Human Action Analysis Based on Improved SlowFast Model Using Action Quantization

Aiming at the problem that the surveillance camera relies on the site monitoring of personnel in the real shopping mall scene, based on the in-depth study of the SlowFast network structure and ResNet network, the SE (Squeeze and Extension) module (SE-RES) is introduced into ResNet. This paper proposes an improved SlowFast network architecture (SE-SlowFast). Using the optimized network structure, action groups are obtained through extensive video analysis using real mall surveillance videos as input sources. This paper proposed a method to quantify the action, and used the action group to analyze the relationship between the actions, so as to further show the action more intuitively in the surveillance video, so as to realize the extraction and analysis of abnormal segments in the surveillance video. This experiment gives the results of quantifying some actions, which is more intuitive compared to the traditional visual observation. This method can provide auxiliary services for surveillance video detection personnel, and greatly improve work efficiency and intelligence level.