【摘要】 提出一种多算法融合的跌倒行为识别算法。首先，针对人体目标的特征对YOLOv3 tiny检测算法进行改进，有效框定人体动态目标区域，提取出目标前景；在此基础上利用AlphaPose姿态识别框架识别出人体骨骼关键点，得到人体主要关节图；最后以人体关节图坐标信息为输入，通过时空图卷积神经网络对跌倒等动作进行检测识别，满足对不同场景跌倒的有效检测。实验结果表明：融合算法改善了不同场景下跌倒行为的检测效果，检测的准确率可达到97.4%,并有效降低了误检率。更多还原

【Abstract】 A fall behavior recognition algorithm based on multi algorithm fusion is proposed. Firstly, Yolov3 tiny detection algorithm is improved to effectively frame the human anchor and extract the target prospect according to the characteristics of human; then alphapose gesture recognition framework is used to identify the key points of human skeleton, and the main joint diagram of human body is obtained; Finally, taking the coordinate information of human joint diagram as input, the spatiotemporal graph convolution neural network is used to detect and identify falls and other actions, which can effectively detect falls in different scenes. The experimental results show that the fusion algorithm can improve the detection effect of fall behavior in different scenes, the detection accuracy can reach 97.4%, and effectively reduce the false detection rate.更多还原