基于改进YOLOv8pose的校园体测运动姿势识别研究  

Research on human motion pose recognition algorithm based on improved YOLOv8pose

作　　者：罗智杰 王泽宇 岑飘 刘文静 郭建军 Luo Zhijie;Wang Zeyu;Cen Piao;Liu Wenjing;Guo Jianjun(Zhongkai University of Agriculture and Engineering College of Information Science and Technology,Guangzhou 510225,China;Zhongkai University of Agriculture and Engineering Intelligent Agriculture Engineering Research Center,Guangzhou 510225,China;Zhongkai University of Agriculture and Engineering Guangzhou Key Laboratory of Agricultural Product Quality Safety Traceability Information Technology,Guangzhou 510225,China)

机构地区：[1]仲恺农业工程学院信息科学与技术学院,广州510225 [2]仲恺农业工程学院智慧农业工程技术研究中心,广州510225 [3]仲恺农业工程学院广州市农产品质量安全溯源信息技术重点实验室,广州510225

出　　处：《电子测量技术》2024年第19期24-33,共10页Electronic Measurement Technology

基　　金：广东省自然科学基金面上项目(2021A1515011605);广东省企业科技特派员项目(GDKTP2021004400);广州市增城区农村科技特派员项目(2021B42121631);2022年仲恺农业工程学院研究生教育创新计划项目(KA220160228)资助。

摘　　要：针对现有运动姿势识别算法在人体姿态检测的准确度和效率上的不足,本文提出一种基于改进YOLOv8pose的高效检测算法。该算法通过引入RL_SEAM模块优化关键点的遮挡情景,结合C2f-Context机制增强上下文信息的利用,提升模型对复杂姿态的识别能力,并利用Pose_SA轻量化检测头提升模型对运动姿势识别的效果与效率。实验结果显示,改进后的YOLOv8pose算法在人体运动姿势识别任务中取得了显著的提升,其参数量及模型大小相比原YOLOv8n基准模型分别降低了14.24%和10.94%,同时精确率、召回率及平均精度均值相较于原模型分别提高了7.60%、7.60%和10.54%。因此,本文提出的YOLOv8-LSP模型有助于解决人体运动姿势识别任务中面临的关键点遮挡、复杂多变姿态等难题。In response to the shortcomings of existing motion pose recognition algorithms in terms of accuracy and efficiency in human pose detection,this paper proposes an efficient detection algorithm based on an improved YOLOv8pose.This algorithm optimizes the occlusion scenario of key points by introducing the RL-SEAM module,enhances the utilization of contextual information through the C2f-Context mechanism,enhances the model′s ability to recognize complex poses,and uses the Pose_SA lightweight detection head to improve the model′s effectiveness and efficiency in recognizing motion poses.The experimental results show that the improved YOLOv8pose algorithm has achieved significant improvement in human motion pose recognition tasks.Its number of parameters and model size have been reduced by 14.24 and 10.94 percentage points respectively compared to the original YOLOv8n benchmark model.At the same time,accuracy,recall,and mean average precision have been improved by 7.60、7.60 and 10.54 percentage points respectively compared to the original model.Therefore,the YOLOv8-LSP model proposed in this article helps to solve the challenges faced in human motion pose recognition tasks,such as key point occlusion and complex and variable postures.

关 键 词：YOLOv8 运动姿势识别 RL_SEAM C2f-Context Pose_SA 

分 类 号：TN98[电子电信—信息与通信工程]