Human Pose Estimation Based On Deep Feature Fusion

Human pose estimation is thse process of obtaining the specific positions of various joints of the human body from image or video information,which has been widely used in human-computer interaction,video surveillance,virtual reality and other fields.Human pose estimation algorithms based on color images are easily influenced by factors such as color and environment,while depth images are more robust for clothing,skin color,and occlusion,especially more adaptive to the challenges of complex environments.This paper mainly studies the human pose estimation method based on depth image,which can effectively solve the problems of noise interference and data redundancy of the depth image,so as to improve the prediction accuracy of the human body pose estimation mode.The main work of the paper includes:(1)Aiming at the problem that depth images contain lots of non-Gaussian noise,a laminar denoising algorithm for depth images is studied.Firstly,the depth level interval of image is determined by the noise intensity and the detection range of depth image.Then median filtering algorithm is used to denoise the depth image of each level,Finally the complete denoised depth image are obtained aftering merging and repairing.The experimental results on the ITOP dataset show our denoising method could effectively eliminate non-Gaussian noise in depth images.(2)Aiming at the problem of insufficient extraction of human body information and redundancy in depth images,a deep image feature fusion model is studied.By extracting the depth comparison feature,BPOF feature and directional gradient feature of the human body in the depth image,and using the feature information entropy and mutual information as evaluation criteria for feature fusion,the human body information representation ability in the depth image is improved and the redundancy of the feature is reduced.The experimental results on the ITOP dataset show our feature fusion method could better extract the key information of the human body and remove the redundancy of deep features.(3)The denoising method and feature fusion model studied in this paper are applied to human pose estimation.Manifold Gaussian Process is used as the regression model to realize the prediction of human joint information.The experimental results carried on ITOP data set show that our model could make full use of the information in the depth image and improve the accuracy of human pose estimation.At the same time,compared with random forest,Support Vector Machine and standard Gaussian Process estimation model,the manifold Gaussian Process regression model achieves better prediction accuracy in the joint parts of the human body with higher flexibility.