摘要
针对摔倒检测难、检测精度低、误检率高等问题提出了一种基于骨骼姿态关键点和卷积神经网络的摔倒检测算法。该算法通过OpenPose对连续n帧中的运动目标进行关键点检测,以VGG预训练网络作为骨架,对运动目标进行姿态特征提取,并将所提取的姿态特征以支持向量机的方法进行分类实验,有效区分坐、躺、蹲等与摔倒相似的行为。测试所使用的数据集包括一系列自建摔倒视频并结合包括走、蹲、躺、坐、跳等五种非摔倒行为。检测结果的灵敏度为96.52%,特异性为96.37%,相比同类检测算法有较大提升。
Aiming at the problems of difficult fall detection,low detection accuracy and high false detection rate,a fall detection algorithm based on the key points of skeletal posture and convolutional neural network is proposed.This algorithm uses OpenPose to detect key points of moving targets in consecutive frames.The system uses the VGG pre-training network is used as the backbone.It extracts the pose features of the moving targets,and the extracted pose features are classified by the support vector machine method,which effectively distinguishes the behaviors that be similar to the fall.The data set used in the test consists of a series of self-built fall videos combined with five non-falling behaviors including walking,squatting,lying,sitting,and jumping.The sensitivity of the detection result is 96.52%,and the specificity is 96.37%,which get a greatly improvement compared with the similar detection algorithms.
作者
尹志成
徐熙平
孙也尧
YIN Zhi-cheng;XU Xi-ping;SUN Ye-yao(School of Optoelectronic Engineering,Changchun University of Science and Technology,Changchun 130022)
出处
《长春理工大学学报（自然科学版）》
2021年第3期15-21,共7页
Journal of Changchun University of Science and Technology(Natural Science Edition)
基金
吉林省科技发展计划项目(20170204048GX)。
作者简介
尹志成(1991-),男,硕士研究生,E-mail:124434889@qq.com;通讯作者:徐熙平(1969-),男,博士,教授,博士生导师,E-mail:xxp@cust.edu.com。
