Research on pedestrian fall detection method based on deep learning

Fall detection technology based on deep learning can provide an emergency rescue window for high-risk groups by identifying fall behaviors in real time, which has important social application value. In order to cope with the challenges of complex background interference, high real-time detection requirements, strict accuracy requirements and the need to control hardware costs in actual scenarios, this paper proposes an improved YOLOv5 detection model, introduces the SENet channel attention mechanism, replaces the original SPPF module with the SPPFCSPC structure, and adopts the BiFPN bidirectional weighted feature pyramid. Through the learnable cross-scale feature weight allocation, it solves the conflict of deep and shallow layer feature transmission and enhances the ability to capture small fall actions. The model of this study is compared with other schemes. The experimental results show that the improved scheme of this study can effectively improve the performance of the model. Experimental data show that the [email protected] of this research scheme reaches 78.9%, which is 3.5% higher than the original YOLOv5s. While maintaining real-time performance, it significantly meets the high-precision requirements of fall detection and provides a reliable technical solution for the monitoring of high-risk groups.