基于 STRS⁃YOLO和OpenPose的跌倒检测算法研究

打开文本图片集

关键词：跌倒检测；STRS⁃YOLO；OpenPose；SA⁃Net；注意力机制；姿态识别；重参数化中图分类号：TN911.23⁃34 文献标识码： A 文章编号：1004⁃373X（2025）12⁃0129⁃08

Abstract：The increasing aging of society has made falling accidents become a prominent health threat for the elderly population. Accurately detecting falling behavior is crucial for preventing elderly people from getting hurt and improving their quality of life. The current fall detection technology is affected by environmental interference （such as changes in lighting， complex backgrounds， etc.）， resulting in insufficient recognition accuracy. In allusion to the limitations of traditional methods in capturing complex details， optimizing fall detection conditions， and detecting occluded scenes， a fall detection algorithm that integrates STRS ⁃ YOLO and improved OpenPose is proposed to achieve the detection and classification of fall behavior. In the backbone network， Swin Transformer module is used to instead of C3 module， and in the neck network， reparameterized RefConv refocus convolution is used to instead of traditional convolution， constructing a channel space mixed attention mechanism （SA Net） model. The DenseNet module is used to connect each layer directly with all previous layers in the channel dimension， and the key points of a single human target are detected by means of improved OpenPose network. A multi⁃parameter fall judgment strategy is set up to optimize the classification and regression performance. The experimental results show that the proposed fall detection model can achieve high accuracy and intelligent detection of falls in the elderly， which has strong research value and broad application prospects.

Keywords：fall detection; STRS⁃YOLO; OpenPose; SA⁃Net; attention mechanism; posture recognition; reparameterization

0 引 言

老龄化社会给医疗保健系统带来了前所未有的压力，医疗保健的需求正迅速增长，而现有的医疗资源和机构护理能力却难以满足这一庞大需求，导致医疗支出激增。（剩余10042字）