基于姿态估计的动作识别方法研究  

Research on Action Recognition Method Based on Pose Estimation

作　　者：江兆银[1] 黄树成[2] 陶哲 JIANG Zhaoyin;HUANG Shucheng;TAO Zhe(Yangzhou Vocational University,225009,Yangzhou,Jiangsu,PRC;Jiangsu University of Science and Technology,212100,Zhenjiang,Jiangsu,PRC)

机构地区：[1]扬州市职业大学,江苏扬州225009 [2]江苏科技大学,江苏镇江212100

出　　处：《江西科学》2023年第3期581-586,613,共7页Jiangxi Science

基　　金：国家自然科学基金资助项目(61772244);江苏省高等教育教改研究项目(2021JSJG161)。

摘　　要：姿态估计和人体动作识别是计算机视觉领域内的两大热点研究方向。人体姿态估计通常面临着坐标点误差等问题。而光照、障碍物遮挡、人体自遮挡以及动作类别混淆等因素的影响会加大动作识别的难度和精度。针对以上问题,现提出如下解决方法:设计一种用Kinect采集人体三维关节点坐标的方法,使用数据投影的方法解决采集过程中出现的坐标点不对应问题,并给出人体姿态特征描述方法。利用卷积神经网络在姿态估计结果的基础之上进行动作识别,使用知识蒸馏的方法建立更为统一且鲁棒的特征空间,在采集到的数据集上获得较高准确率。Pose estimation and human action recognition are two hot research directions in the field of computer vision.Human pose estimation usually faces problems such as coordinate point errors.The influence of factors such as illumination,obstacle occlusion,human self-occlusion,and action category confusion will increase the difficulty and accuracy of action recognition.In view of the above problems,the following solutions are proposed:design a method to collect the coordinates of three-dimensional joint points of the human body with Kinect,use the method of data projection to solve the problem of mismatching coordinate points during the acquisition process,and give a description method of human body posture features.The convolutional neural network is used for action recognition based on the pose estimation results,and the knowledge distillation method is used to establish a more unified and robust feature space,and a higher accuracy rate is obtained on the collected data set.

关 键 词：计算机视觉 姿态估计 动作识别 卷积神经网络 知识蒸馏 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]