基于改进YOLO-Pose轻量模型的多人姿态估计  

Multi-person pose estimation based on an improved lightweight YOLO-Pose model

作　　者：李传江[1] 汪著名 张崇明[1] LI Chuanjiang;WANG Zhuming;ZHANG Chongming(College of Information,Mechanical and Electrical Engineering,Shanghai Normal University,Shanghai 201418,China)

机构地区：[1]上海师范大学信息与机电工程学院,上海201418

出　　处：《上海师范大学学报（自然科学版中英文）》2024年第2期188-194,共7页Journal of Shanghai Normal University(Natural Sciences)

摘　　要：YOLO-Pose作为人体姿态估计算法模型,在精度和速度上有着不错的表现,但其在复杂和有遮挡的场景下存在误检率较大的问题,并且模型的复杂度仍然有优化的空间.针对这几个问题,通过选取Slim-neck模块和Res2Net模块,重新设计其特征融合层,减少其计算量和参数量,提高特征提取能力,在提升精度的同时,使模型轻量化;引入EIoU损失函数,加快边框检测的收敛速度,并提高定位的准确性.在压缩的OC_Human数据集上进行测试,改进后的模型与YOLO-Pose相比,P值、mAP@0.5和mAP@.5:95分别提高了10.6,3.1和2.9个百分点.此外,参数量和计算量也分别减少了16.7%和19.3%,在精度和轻量化方面均有所提升,为其应用在资源有限的边缘计算设备提供了可能性.YOLO-Pose as a human pose estimation algorithm model had a good performance in terms of accuracy and speed,which suffered from a large false detection rate in complex and occluded scenes on the other hand.There was still room for optimization of the model complexity.In this paper,these issues were addressed by incorporating the Slim-neck module and Res2Net module to redesign the feature fusion layer,reducing computational and parameter overhead while enhancing the information extraction capability of feature extraction.Furthermore,the EIoU loss function was introduced to accelerate the convergence speed of bounding box detection and to improve localization accuracy.Experimental results on the compressed OC_Human dataset demonstrated that the improved model achieved a 10.6% improvement in P-value,a 3.1% increase in mAP@0.5,and a 2.9% increase in mAP@.5:95 compared to the original YOLO-Pose model,respectively.Moreover,the amount of parameters(Params)and computational complexity(GFLOPs)were reduced by 16.7% and 19.3%,respectively.The improved model showed enhanced accuracy and lightweight characteristics,which was suitable for deployment on resource-constrained edge computing devices.

关 键 词：人体姿态估计 YOLO-Pose 轻量化 Slim-neck 

分 类 号：TP394.1[自动化与计算机技术—计算机应用技术]