基于计算机视觉的人体姿态识别研究  

Research on Human Posture Recognition Based on Computer Vision

作　　者：盛洋 王健庆[1] SHENG Yang;WANG Jianqing(School of Medical Technology and Information Engineering,Zhejiang Chinese Medical University,Hangzhou 310053,China)

机构地区：[1]浙江中医药大学医学技术与信息工程学院,浙江杭州310053

出　　处：《现代信息科技》2022年第16期87-91,95,共6页Modern Information Technology

基　　金：浙江省基础公益研究计划项目(LGG19F030008);教育部留学回国人员科研启动基金资助项目(2014)。

摘　　要：人工智能交互领域的不断发展,使计算机视觉技术的研究更加深入且多样化。人体姿态识别是计算机视觉中的一个重要的基础问题。文章基于计算机视觉开展人体姿态识别研究,通过VGGNet模型实现神经网络预测,采用Openpose算法获取18个人体关键点位置和置信度,根据PAF和二分图算法实现关节拼接。实验基于MS COCO 2017数据集建立优化模型,在测试集上达到了96.31%的准确率。最后,文章实现了能够识别出人体骨骼姿态的系统。The constant development of the field of Artificial Intelligence interaction makes the research of computer vision technology become deeper and more diverse.Human posture recognition is an important basic problem in computer vision.This paper researches the human posture recognition based on computer vision,realizes the neural network prediction through VGGNet model,obtains the position and confidence level of 18 human key points by using Openpose algorithm,and realizes the joint splicing according to the PAF and bipartite graph algorithm.The experiment establishes optimization model based on MS COCO 2017 data set,and it achieves 96.31% accuracy rate on the test set.Finally,the paper realizes a system that can recognize the posture of human bones.

关 键 词：人体姿态识别 姿态检测 Openpose 神经网络 可视化 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]