3D Human Pose Estimation Based On Deep Learning In Home Healthcare Scene

With the development of society,people pay more and more attention to personal health,and the most concerned is home healthcare.The significance of home healthcare is very significant for specific groups,such as the elderly,children and pregnant women.Human pose estimation is one of the important technologies for home healthcare system,which is indispensable component for the application of action recognition,interactive operation and danger warning.Accurate and rapid human pose estimation is an indispensable technology for the home healthcare system.In the real world,there are several reasons influence the accuracy of human pose estimation in three-dimensional space,such as the change of camera angle,the color of clothing and the ambiguity of spatial position.Researchers proposed various methods to solve these problems for higher accuracy,such as multi-view based method and depth information based method,but these methods still have many limitations,for instance,the cost and scene constraint.After the emergence of deep learning,the 3D pose estimation method based on monocular images has become a new research area,however,this kind of method still could not solve well the 3D human pose estimation problem in home healthcare scene.In this study,we use a top-down approach to solve the 3D human pose estimation problem in the home healthcare scene.The method is divided into two stages,we use target detection algorithm as human body detector to detect single-person in the image at the first stage.The second stage is single person 3D pose estimation,the network not only detect the 2D information but also needs to restore the depth information of each joint.Considering algorithm generalization in home healthcare scene,we use the midpoint of the crotch as the root node and restore the 3D human pose through the relative relationship between the joint and root with spatial context information.The single-person 3D pose estimation framework is stacked network structure,we use the hourglass module as the base module to learn spatial context information,multi-scale information fusion and context information inference can better learn related information betweent joints.We adopt the residual learning idea and regularization to solve problems caused by the depth network,such as under-fitting,gradient dispersion and difficult training,and the intermediate supervision is adopted in the model training.We present experiments on the Human3.6m and HumanEva datasets to verify the accuracy of single person 3D pose estimation,and we carried out experimental analysis on the actual home health care scene to verify the feasibility of the overall algorithm.