Study Of 3D Human Pose Estimation Based On Depth Image

Human pose estimation is an important research hotspot in the field of computer vision,which can be prevalently applied to video surveillance,human motion recognition and human-machine.According to the different types of used images,it can be divided into two major fields,based on color image and depth image.Among them,the depth image only records the distance information between the camera and the target,which is color-independent,barely interfered by environmental factors such as illumination changes,and has irreplaceable advantages in the protection of privacy.However,due to the simplicity of depth image pose and the poor performance of existing pose estimation algorithms,the 3D human pose estimation of depth image is still a difficult and hot research topic at present.In this paper,in order to achieve 3D human pose estimation of depth images,the random forest method and convolutional neural network method based on deep learning are studied.In the field of pose estimation based on random forest,limited by the training samples,this paper only studies the upper limb pose estimation method.As far as human pose estimation method of deep learning is concerned,there is no limitation of the above research,so the human pose estimation based on deep learning method of the whole human body is studied.The research contents are as follows:1.This paper propose an improve method which can reduce the misclassification in human pose estimation based on random forest and increase the accuracy,including adaptive fusion feature extraction and misclassification processing mechanism.Firstly,we improve the method of feature extraction to adaptively extract deep fusion feature,so that,both distance information and part information could enhance feature expression.Furthermore,owing to inspiration from error cluster analysis and iteration thought,the misclassification processing mechanism is proposed to handle misclassification appearance.Finally,we achieve accurate human pose estimation from single depth image by applying the principal direction vector based on the improve principal direction analysis algorithm.2.In this paper,3D human pose estimation method of depth images based on stacked hourglass networks is proposed with sufficient joint point annotation.The propose network structure is composed of 2D regression module and depth module.It can complete the training of 3D regression model based on depth image through strong supervised learning technology and end-to-end training mode,so as to achieve 3D human pose estimation of depth image.3.Finally,we focus on the lack of depth labels in depth images and the low generalization ability of models due to simple gestures,this paper innovatively propose a 3D human pose estimation method of multi-source images based on weakly-supervised approach.This method mainly includes the following points.(1)Using multi-source image fusion training method to improve the generalization ability.(2)Proposing weakly-supervised approach to solve the problem of label insufficiency.(3)In order to improve the attitude estimation results,this paper improve the residual module.Finally,we achieve accurate 3D human pose estimation from single depth images.