基于AlphaPose模型的远距离行人头部姿态估计算法  

Algorithm for long-range pedestrian head pose estimationbased on AlphaPose model

作　　者：赵思源 彭春蕾 张云 刘海涛 胡瑞敏 ZHAO Si-yuan;PENG Chun-lei;ZHANG Yun;LIU Hai-tao;HU Rui-min(School of Cyber Engineering,Xidian University,Xi′an 710071,China;Artificial Intelligence and Big Data Applications Research Institution,Nanning 530000,China)

机构地区：[1]西安电子科技大学网络与信息安全学院,陕西西安710071 [2]广西产研院人工智能与大数据应用研究所有限公司,广西南宁530000

出　　处：《陕西科技大学学报》2023年第2期191-198,共8页Journal of Shaanxi University of Science & Technology

基　　金：国家自然科学基金项目(U22A2035,62276198);广西自然科学基金重点项目(粤桂联合基金项目)(2021GXNSFDA075011)。

摘　　要：在公共场所的监控视频中,远距离的行人目标头部区域占比往往较小并且头部区域分辨率较低,仅靠目标检测算法和头部姿态估计模型检测并分类头部特征来判定头部姿态或注视方向存在较大难度.考虑到目前不同分辨率下的人体骨骼关键点估计算法越来越成熟,本研究提出一种利用骨骼关键点和头部姿态之间的关系来进行远距离行人的头部姿态估计算法.该算法首先利用AlphaPose模型检测出二维人体骨骼关键点的全局坐标值,然后根据部分骨骼关键点的坐标值计算出头部朝向象限角度,最后根据预设角度范围计算并可视化出视线落点.目前,尚没有公开发布的可用于远距离行人头部姿态估计的数据集,因此本文建立了一个远距离行人头部姿态系统,以对现有的数据集进行标注,并利用标注的数据集对提出的方法进行测试.结果表明,本文提出的算法可以较精准地判定到远距离行人的头部姿态和注视方向,准确率达到69.7%.Pedestrian targets outside the visible distance in surveillance videos of public places have smaller heads and lower head area resolution,it is difficult to determine the head posture or gaze direction only by the target detection algorithm and the head posture estimation model to detect and classify the head features.Considering that the current human skeleton key point estimation algorithms at different resolutions are becoming more and more mature,this paper proposes a head pose estimation algorithm based on the relationship between skeleton key points and head posture for long-distance pedestrian head pose estimation.The algorithm first uses the AlphPose model to detect the global coordinate values of 2D human skeleton key points,then calculates the quadrant angle of the head according to the coordinate values of some skeleton key points,and finally calculates and visualizes the sightline according to the preset angle range.At present,there is no publicly released data set that can be used for long distance pedestrian head pose estimation,so we built a long-distance head pose estimation system to annotate existing data sets.After testing the labeled data sets,the results show that the proposed algorithm can more accurately determine the long-distance pedestrian head pose estimation and gaze direction,with an accuracy rate of 69.7%.

关 键 词：远距离行人头部姿态估计 注视方向估计 人体骨骼关键点检测 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]