基于改进YOLOv8n-Pose的人体关键点检测算法  

Human Key Point Detection Algorithm Based on Improved YOLOv8n-Pose

作　　者：雷亮 陈毅 刘学涵 赵锦 陈小庆 周华勇 徐山雯 LEI Liang;CHEN Yi;LIU Xuehan;ZHAO Jin;CHEN Xiaoqing;ZHOU Huayong;XU Shanwen(School of Computer Science and Engineering,Chongqing University of Science and Technology,Chongqing 401331,China;School of Big Data and Information Industry,Chongqing City Management College,Chongqing 401331,China)

机构地区：[1]重庆科技大学计算机科学与工程学院,重庆401331 [2]重庆城市管理职业学院大数据与信息产业学院,重庆401331

出　　处：《重庆科技大学学报(自然科学版)》2025年第1期83-89,共7页Journal of Chongqing University of Science and Technology(Natural Sciences Edition)

基　　金：重庆市教委科学技术研究项目“基于应急决策的校园社区大数据平台研究与应用”(KJQN202203314);“基于YOLO-Pose姿态特征检测融合时序的独居老人跌倒识别算法研究”(KJQN202303305);“基于深度学习的学生课堂行为检测方法研究”(KJQN202403322);2021年重庆市属本科高校与中科院所属院所合作项目“工业互联网内生安全关键技术研究与协同创新”(HZ2021015)。

摘　　要：针对YOLOv8n-Pose模型在边缘设备部署面临的计算资源和存储空间受限问题,提出了一种基于改进YOLOv8n-Pose的人体关键点检测算法。设计了一种基于上下文导引特征融合模块,以提升模型的特征表达能力和检测精度,并引入深度卷积进行图像特征处理,使得检测头的参数量更少、计算量更低。在COCO-Pose数据集上的实验结果表明,改进后的YOLOv8n-Pose模型在保证检测精度提升的同时,使得参数量降低了19.5%、计算量减少了32.6%,从而提高了模型在边缘场景中的适用性。Aiming at the limitations of computational resource and storage space,which the YOLOv8n-Pose model faces when deployed on edge computing devices,this paper proposes an improved human key point detection algorithm based on the improved YOLOv8n-Pose model.A context-guided feature fusion model is designed to improve the expression of CGFF model′s feature and detection accuracy.Meanwhile,the depthwise convolution is introduced to improve the head part of the model,significantly reducing the number of parameters and calculations.The experimental results of the COCO-Pose dataset show that the optimized YOLOv8n-Pose model reduces the overall network parameters by 19.5%and the amount of calculation by 32.6%while ensuring the improvement of detection accuracy,which improves its applicability in edge computing scenarios.

关 键 词：YOLOv8n-Pose算法 深度卷积 上下文导引特征融合 姿态估计 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]