基于YOLOv8-pose的人体姿态检测模型  

Human pose detection model based on YOLOv8-pose

作　　者：方晓柯 黄俊[1] FANG Xiaoke;HUANG Jun(School of Communications and Information Engineering,Chongqing University of Posts and Telecommunications,Chongqing 400065,China)

机构地区：[1]重庆邮电大学通信与信息工程学院,重庆400065

出　　处：《激光杂志》2025年第3期50-57,共8页Laser Journal

基　　金：国家自然科学基金(No.61771085)。

摘　　要：针对多人人体姿态估计场景下关节点检测丢失以及小目标无法识别等问题,提出了一种改进的YOLOv8-Pose模型。该算法的核心改进在于使用可变性卷积DCNV2替换了C2F模块中的卷积,从而增强了网络的特征提取能力。同时,使用加权双向金字塔BiFPN模块替换原模型中的特征融合模块,保留小目标信息的同时,融合更多的浅层信息,以提高识别准确度。最后,为了进一步加强对关键部位的捕获和分析能力,引入了SimAM注意力机制,对局部特征进行加权处理。实验结果表明,在CrowdPose数据集上,该算法的检测精度达到了74.5%,比原模型高出了3.3%。与原YOLOv8-pose模型相比,改进后的模型不仅具有更高的检测精度,而且在小目标的识别效果上也有显著的提升。由此可见,改进后的网络能更加精确、有效地应用于多人人体姿态检测。In order to solve the problems of loss of joint point detection and inability to identify small targets in the scenario of multi-person human pose estimation,an improved YOLOv8-Pose model was proposed.The core improvement of the algorithm is that the convolution in the C2F module is replaced by the variable convolution DCNV2,which enhances the feature extraction ability of the network.At the same time,the weighted bidirectional pyramid BiFPN module is used to replace the feature fusion module in the original model,which aims to retain the small target information and fuse more shallow information to improve the recognition accuracy.Finally,in order to further strengthen the ability to capture and analyze key parts,the SimAM attention mechanism was introduced to weight the local features.Experimental results show that the detection accuracy of the algorithm reaches 70.5% on the CrowdPose dataset,which is 3.3% higher than that of the original model.Compared with the original YOLOv8-Pose model,the improved model not only has higher detection accuracy,but also has a significant improvement in the recognition effect of small targets.It can be seen that the improved network can be applied to multi-person human posture detection more accurately and effectively.

关 键 词：姿态识别 关节点检测 YOLOv8-Pose DCNV2 SimAM 

分 类 号：TN209[电子电信—物理电子学]