基于卷积注意力模块的人体姿态估计研究

打开文本图片集

中图分类号：TP391.4 文献标志码：A

DOI：10.13338/j. issn. 1674-649x.2025.02.001

Research on human pose estimation based on convolutional block attention module

LIAN Jihong，XUE Weige，WANG Yannian，ZHANG Nan （School of Electronics and Information，Xi'an Polytechnic University，Xi'an 71O048，China）

Abstract In order to further ensure pedestrian safety in autonomous driving， this article addresses the issues of key point false detection，missed detection，and redundancy in human pose estimation.HRNet was used as the backbone network for algorithm optimization to further improve model detection accuracy.Firstly， a human pose estimation model inference network RSGNet was designed for image feature extraction，which eliminates the influence of interfering keypoints during the keypoint inference process and improves the effective utilization of keypoint information by the model. Secondly， in response to the problem of incomplete image detail information caused by self occlusion or external interference，a convolutional block attention module （CBAM） was added to image feature processing. This module combines spatial and channel correlation fusion information to reduce the negative impact of foreground，background，and other information on image processing. The experimental results show that compared with the benchmark model HRNet method， the improved network model significantly improves the detection accuracy of human pose estimation，with an average precision（AP） increase of 7.3% in the public dataset COCO，and the AP in the public data set MPII is increased by 3.0%

Keywordspose estimation;autonomous driving;key points;attention mechanisms;spatial attention;channelattention

0引言

随着机器视觉的迅速发展，人体姿态估计技术作为解决人体关键点（如头、肩、肘等）在图像、视频中的位置表示方法[1]，该技术旨在对人体关节点的位置进行检测和定位，可以实时获取人体各部位的运动信息，从而输出全部或局部关节的信息。（剩余14093字）