基于改进YOLOv7-Pose网络综合体测计数算法  

Comprehensive Body Measurement Counting Algorithm Based on Improved YOLOv7-Pose Network

作　　者：孟丹 单巍[1,2] 慕灯聪 李峥 MENG Dan;SHAN Wei;MU Dengcong;LI Zheng(School of Physics and Electronic Information,Huaibei Normal University,235000,Huaibei,Anhui,China;Anhui Province Key Laboratory of Intelligent Computing and Applications,235000,Huaibei,Anhui,China)

机构地区：[1]淮北师范大学物理与电子信息学院,安徽淮北235000 [2]智能计算及应用安徽省重点实验室,安徽淮北235000

出　　处：《淮北师范大学学报（自然科学版）》2024年第4期73-78,共6页Journal of Huaibei Normal University：Natural Sciences

基　　金：安徽省高校自然科学研究重点项目(KJ2020A0027,2022AH050392);安徽省新时代育人质量工程项目(2023lhpysfjd044);安徽省教育厅质量工程教学研究项目(2022jyxm1387)。

摘　　要：针对体测项目中引体向上和仰卧起坐计数效率及准确率低问题,设计基于改进YOLOv7-Pose网络综合体测计数算法。该算法通过检测人体关键点数据来计算肢体关节角度,设定阈值判断动作是否达标并实现计数。针对应用场景,将网络检测重点放在人体姿态中大目标上,删除小目标感受野模块和面部关键点,减小模型规模,提高实时性,以便移植到便携设备中;在原始YOLO7-Pose网络中引入SE注意力机制模块,提高模型特征提取能力;为最小化预测边界框和真实边界框的高度和宽度差异,使网络有更快收敛速度和更好定位结果,用EIoU替换原网络中的CIoU损失函数。在COCO数据集上实验结果表明,改进后的算法模型mAP@0.5提高了0.9个百分点,FPS显著提升,计数准确率均达到98%以上,满足体测场景对计数准确率的要求。Aiming at the low counting efficiency and accuracy of pull-ups and sit-ups,a comprehensive counting algorithm based on improved YOLOv7-Pose network was designed.The algorithm calculates the limb joint angle by detecting the key point data of the human body,sets the threshold to judge whether the action reaches the standard and realizes the counting,according to the application scenario,the network detection focuses on the large target of human posture,deletes the small target receptive field module and facial key pointes,reduces the model size,improves the real-time performance,and can be transplanted into portable devices.The SE attention mechanism module is introduced into the original YOLO7-Pose network to improve the feature extraction ability of the model.In order to minimize the difference between the height and width of the predicted bounding box and the real bounding box,the CIoU loss function in the original network is replaced by EIoU in order to achieve faster convergence and better positioning results.Experimental results on COCO dataset show that the improved algorithm model mAP@0.5 increases by 0.9 percentage points,FPS is significantly improved,and the counting accuracy rate reaches above 98%,which meets the requirement of counting accuracy in physical test scenarios.

关 键 词：人体关键点检测 SE注意力机制 引体向上计数 仰卧起坐计数 YOLOv7-Pose 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]