文献类型：期刊文献

中文题名：基于深度学习的人体姿态估计与追踪

英文题名：Human Pose Estimation and Tracking Based on Deep Learning

作者：张雪芹[1];朱荟潼[1];王宁[1]

机构：[1]华东理工大学信息科学与工程学院,上海200237

年份：2024

卷号：50

期号：5

起止页码：750

中文期刊名：华东理工大学学报（自然科学版）

外文期刊名：Journal of East China University of Science and Technology

收录：;Scopus;北大核心:【北大核心2023】；CSCD:【CSCD_E2023_2024】；

基金：国家自然科学基金(51975213)。

语种：中文

中文关键词：人体姿态估计;AlphaPose;YOLOv5s;BCNet;SeqNet

外文关键词：human pose estimation;AlphaPose;YOLOv5s;BCNet;SeqNet

摘要：随着深度学习技术的发展,基于卷积神经网络的人体姿态估计和追踪的准确率得到大幅提高。但在面对遮挡问题时,还存在人体关键点检测困难、姿态追踪精度偏低和速度较慢等问题。本文针对这些问题,构建了一个ybasTrack多人姿态估计和追踪模型;提出采用一种改进的YOLOv5s网络进行目标检测;采用BCNet分割网络区分遮挡与被遮挡人体,限定人体关键点定位区域;基于Alphapose的SPPE(Single-Person Pose Estimator)进行改进,优化人体关键点检测结果;采用改进的Y-SeqNet网络进行行人重识别,采用MSIM(Multi-Phase Identity Matching)身份特征匹配算法对人体框、人体姿态和人体身份信息进行匹配,实现人体姿态追踪。实验表明,所提算法对遮挡场景下的人体姿态估计和姿态追踪具有较好的效果,模型运行具有较快速度。
With the development of deep learning technology,the accuracy of human posture estimation and tracking based on convolutional neural network has been significantly improved.However,when facing occlusion problems,there are still difficulties in detecting the key points of the human body,low posture tracking accuracy,and slow speed.In this paper,a ybasTrack multi-person pose estimation and tracking model is constructed to address these problems.An improved YOLOv5s network is proposed for target detection;a BCNet segmentation network is used to distinguish between occluded and occluded human bodies and limit the localization area of human body key points.Alphapose-based SPPE is improved to optimize the detection results of human key points.An improved Y-SeqNet network is used for pedestrian re-identification,and the MSIM identity feature matching algorithm is used to match the human body frame,posture,and identity information,achieving human body posture tracking.It is shown from experiment results that the proposed algorithm has better performance in human posture estimation and tracking in occlusion scenes,and the model runs at a faster speed.