摘要
Abstract
In order to efficiently screen the key information of abnormal human motions and realize multi-perspective recognition of abnormal human motions,a multi-perspective recognition method for abnormal human motions under keyframe localization is proposed.In this method,a global and partitioned spatial model of the human skeleton sequence is constructed.The spatial characteristics of human motions are described accurately by 3D coordinates of the joint points and formulas for adjacent frame position changes.Then,on the basis of the characteristics of human motions,the velocity and acceleration of joint points are calculated,and threshold filtering is set to identify abnormal keyframes,so as to achieve the localization of human abnormal motion keyframes.The human motion video image frame sequence after the localization of abnormal motion keyframes is input into the LSTM model embedded in the 3D residual network.In this model,a 3D residual network is used to capture the multi-perspective spatiotemporal local features of human abnormal motions,and then the LSTM neural network model is used to forget and screen the multi-perspective spatiotemporal local features,and output the multi-perspective recognition results.The experimental results show that the proposed method can accurately obtain the spatial variation information of joint points,providing a key foundation for multi-perspective recognition.It can effectively locate abnormal motion keyframes within the sequence of human motion video images,and recognize human abnormal motions in multi-perspectives,with significant application effects.关键词
关键帧定位/人体异常动作/多视角识别/LSTM模型/人体骨架/残差网络/局部特征/关节点Key words
keyframe localization/abnormal human motion/multi-perspective recognition/LSTM model/human skeleton/residual network/local feature/joint point分类
信息技术与安全科学
