基于计算机视觉技术的人体姿态跟踪方法  

Human Pose Tracking Method Based on Computer Vision Technology

作　　者：刘恒 高新燕 苏新新[2] 李成龙 LIU Heng;GAO Xinyan;SU Xinxin;LI Chenglong(School of Computer Science and Technology,Shandong Jianzhu University;Shandong Hoteam Software Co.,Ltd.,Ji′nan 250101,China)

机构地区：[1]山东建筑大学计算机科学与技术学院 [2]山东山大华天软件有限公司,山东济南250101

出　　处：《软件导刊》2025年第4期136-146,共11页Software Guide

基　　金：国家自然科学基金项目(62102235);山东省自然科学基金项目(ZR2020QF029);山东建筑大学博士基金项目(XNBS1811)。

摘　　要：人体姿态跟踪在人体运动分析、智能监控、虚拟现实等应用场景中得到了广泛应用。为解决在多人姿态跟踪过程中因多人遮挡、复杂环境等原因导致的目标易丢失、姿态估计不准确等问题,提出一种结合DeepSORT和OpenPose的人体姿态跟踪方法。选择合适的数据集分别训练人体运动跟踪模型和人体姿态估计模型。在人体运动跟踪阶段,基于改进的YOLOv5与改进的DeepSORT算法进行准确地识别与跟踪,减小了模型内存访问成本,引入模型剪枝方法进一步提高了跟踪速度。在人体姿态估计阶段,基于跟踪阶段获得的准确目标,利用OpenPose算法实现了目标人体的姿态估计。相较于对比方法,改进算法在COCO2017数据集上的FLOPs为4.46 G,检测速度大幅提升,在MOT16数据集、PoseTrack2017数据集上的运行时间有较大优势,分别为85.55 s、81.79 s。Human pose tracking has been widely used in application scenarios such as human motion analysis,intelligent monitoring,and vir‐tual reality.To address the issues of target loss and inaccurate pose estimation caused by multiple occlusion and complex environments in the process of multi person pose tracking,the article proposes a human pose tracking method that combines DeepSORT and OpenPose.Select ap‐propriate datasets to train human motion tracking models and human pose estimation models separately.In the human motion tracking stage,based on the improved YOLOv5 and the improved DeepSORT algorithms,accurate recognition and tracking are carried out,reduced the mod‐el memory access cost,and model pruning methods are introduced to further improve tracking speed.In the human pose estimation stage,based on the accurate target obtained in the tracking stage,the OpenPose algorithm is used to achieve the pose estimation of the target human body.Compared with the comparison method,the improved algorithm has FLOPs of 4.46 G on COCO2017 dataset,greatly improving the de‐tection speed,and has a greater advantage in the running time on MOT16 dataset and PoseTrack2017 dataset,which are 85.55 s and 81.79 s respectively.

关 键 词：人体姿态跟踪 目标检测 YOLOv5 DeepSORT OpenPose 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]