基于STGCN算法的视频图像人体动作轮廓动态识别  

STGCN algorithm based dynamic recognition of human motion contour in video image

作　　者：张宗[1] 石林[1] ZHANG Zong;SHI Lin(Changzhou University,Changzhou 213164,China)

机构地区：[1]常州大学,江苏常州213164

出　　处：《现代电子技术》2024年第18期144-148,共5页Modern Electronics Technique

基　　金：中石油科技计划项目(2021DQ06)。

摘　　要：人体动作轮廓在视频中的呈现具有多样性和连续性。人体动作不仅涉及到时间上的变化,还包括空间上的位置关系,受其姿势、速度、方向等影响。人体动作时空信息之间的关联难以充分捕捉,导致动作轮廓识别精度较低。为此,引入时空图卷积网络(STGCN)算法,提出一种视频图像人体动作轮廓动态识别方法。文中采用OpenPose模型从视频图像中提取描述关节点位置的置信图和描述人体关节间连接情况的二维矢量场,构建人体动作骨架图。结合视频帧时间序列组建人体动作骨架时空图,将其作为STGCN模型的输入,通过时空图卷积操作充分捕捉人体动作的时空特征后,采用Softmax层获取动态识别到的视频图像人体动作轮廓;并在STGCN模型中引入两种注意力模块,强化网络特征提取能力,提高动作轮廓识别精度。实验结果表明,所提方法可以有效实现视频图像人体动作轮廓的动态识别,引入的两种注意力模块对STGCN模型进行改进,可提升其动作轮廓识别效果。The presentation of human motion contour in video has diversity and continuity.Human motion not only involve changes in time,but also include position relations in space.Due to the influence of posture,speed,direction,etc.,the correlation between spatio-temporal information of human motion is difficult to fully capture,resulting in low accuracy of the recognition of motion contour.Therefore,the spatio-temporal graph convolutional network(STGCN)algorithm is introduced,and a method of dynamic recognition for human motion contour in video image is proposed.The OpenPose model is used to extract the confidence graph describing the position of the joint points and the two-dimensional vector field describing the connection between human joints from the video image to build the human motion skeleton graph.In combination with video frame time series,the spatio-temporal graph of human motion skeleton is constructed,which is used as the input of STGCN model.After the spatio-temporal features of human motion are fully captured by the convolution operation of the spatio-temporal graph,the dynamic human motion contours in video images are obtained by means of Softmax layer.Two kinds of attention modules are introduced into STGCN model to strengthen the ability of extracting network feature and improve the accuracy of recognizing motion contour.The experimental results show that the proposed method can effectively realize the dynamic recognition of human motion contour in video images.Two attention modules are introduced to improve the STGCN model,which can improve the recognition effect of human motion contour.

关 键 词：时空图卷积网络算法 视频图像 人体动作轮廓 动态识别 注意力机制 骨架图 人体关节点 

分 类 号：TN919.8-34[电子电信—通信与信息系统] TP391[电子电信—信息与通信工程]