A Lightweight Video-Based Behavior Recognition Network Combining MobileNet and Temporal Shift Module

Video action recognition technology has recently gained widespread application in scenarios such as security surveillance, intelligent teaching, and human-computer interaction. However, existing deep learning-based methods face a practical dilemma: while achieving high recognition accuracy, they often suffer from high model complexity, placing heavy demands on computational resources, and hindering efficient operation on edge devices. To address this technical bottleneck, this study innovatively combines a MobileNet backbone with a temporal shift module (TSM) to construct a lightweight action recognition framework. Specifically, this framework leverages MobileNet to extract spatial features from video frames and embeds a TSM module to model temporal features across frames, significantly improving temporal modeling capabilities with virtually no increase in parameter count. Experimental validation on representative datasets such as Hockey Fight and Violent Crowd demonstrates that the new model strikes a good balance between computational efficiency and recognition accuracy, achieving performance comparable to mainstream models and making it particularly suitable for resource-constrained real-world applications.