结合改进Alphapose和GCN的人体摔倒检测模型研究  

A Combination of Improved Alphapose and GCN Human Fall Detection Model

作　　者：曲英伟[1] 梁炜 QU Yingwei;LIANG Wei(School of Software,Dalian Jiaotong University,Dalian 116028,China)

机构地区：[1]大连交通大学软件学院,辽宁大连116028

出　　处：《大连交通大学学报》2023年第2期93-100,共8页Journal of Dalian Jiaotong University

基　　金：国家自然科学基金资助项目(01771087)。

摘　　要：为改进当前人体摔倒检测方法场景适应能力弱、易误检等不足,提出了一种基于人体骨骼关键点和GCN结合的人体摔倒检测模型。在CrownHuman、COCO2017、Le2i等数据集上进行对比试验,试验结果表明优化后的YOLOX人体目标检测算法的平均准确率达到了50.66%,较YOLOv3、YOLOv5提高了9.83%和3.97%。人体姿态估计算法的平均准确率达到了71.6%,优于OpenPose、Mask-RCNN等方法。基于图卷积的人体摔倒检测算法准确率达到92.2%,高于YOLOv5-S+pose等方法。一系列的试验结果表明,所提出的摔倒检测方法具有较高的检测精度。To improve the shortcomings of current human fall detection methods such as weak scene adaptation and easy false detection,a human fall detection model is proposed based on the combination of human skeletal key points and GCN.Comparative experiments are conducted on CrownHuman,COCO2017,Le2i and other datasets.The optimized YOLOX human target detection algorithm achieves an average accuracy of 50.66%,which is 9.83%and 3.97%better than Y0L0v3 and Y0L0v5.The average accuracy of human pose estimation algorithm reaches 71.6%,which is better than OpenPose and Mask-RCNN methods.The accuracy of the graph convolution-based human fall detection algorithm in this paper reaches 92.2%,which is higher than methods such as Y0L0v5-S+pose.A series of experiment results show that the proposed fall detection method has high detection accuracy.

关 键 词：人体摔倒检测 YOLOX Alphapose 人体骨骼关键点 GCN 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP183[自动化与计算机技术—计算机科学与技术]