基于关节点提取的多特征融合人体摔倒检测

Multi-feature Fusion Human Fall Detection based on Joint-point Extraction

意外摔倒是威胁老年人安全的重要因素,能实时高效识别摔倒动作的检测系统可以帮助老年人最大限度地减少摔倒带来的伤害。文章提出一种利用关节点特征结合运动学特征的人体摔倒检测方法。首先使用深度卷积神经网络的人体目标检测算法获取视频中人体的所在位置;然后使用人体姿态估计算法对目标人体进行骨骼关键点提取;最后使用运动学特征人体外接矩形宽高比、质心节点的下降速度、头部关节点与地面之间的距离及人体主躯干、左右腿、左右胳膊和地面之间的夹角与提取的关节点特征进行融合,结果表明摔倒测试在灵敏性、特异性、准确性上分别达到了97.5%、95%、96%的效果。

Accidental falls have become an important factor threatening the safety of the elderly.Detection systems that efficiently identify falls in real time can effectively minimize the injuries caused by falls.This paper proposes a human fall detection method that uses joint-point features combined with kinematic features.Firstly,human object detection algorithm based on depth convolution neural network is used to obtain the location of the human skeleton in the video.The Open Pose algorithm is then used to extract bone key points from the target human body.Then the kinematics features such as the width height ratio of the human body's external rectangle,the falling speed of the centroid node,the distance between the head joint point and the ground,and the included angle between the human body's main trunk,left and right legs,left and right arms and the ground are used to fuse with the extracted joint-point features.The results show that the sensitivity,the specificity and the accuracy have reached 97.5%,95%,and96%respectively in the fall test.