Accurately Estimating The Whole Body Shape And Pose Of Human Body Based On Kinect Single Shot Data

There are many limitations and difficulties in estimating the non-rigid human body under the loose clothes.The occlusion of the clothes and the self-occlusion of the body make the problem even more difficult.Some methods use multi-camera or multi-view data to get more body information,which may be convenient in some specific scenes.Some methods take a simple way to estimate with a single RGB image,but multi-device and source data processing is complex,or simple data does not provide enough information to generate the poor results.In addition,most existing methods do not accurately restore the head and hand status information of the human body.Therefore,this thesis proposes a method of estimating the complete human body in any pose and wearing any dressing by a single-shot RGB-D data with a depth camera.Firstly,the background of this study,as well as status quo of the methods related to human estimation are described,according to their data-obtain way.We also details the static statistical human body model used in this article,and the methods or other related work used in our data processing,and the recent classic methods used for comparison with our methods.Secondly,for overcoming the limitations of the current human body estimation algorithm when recovering from loosely clothed human being,we propose a method for accurately estimating the full human body shape and posture based on the single shot data of the Kinect depth camera.It can accurately estimate 3D model of the corresponding object,and accurately restore the posture and state of the body,the head and the hand.Our method firstly takes a single shot of the current human body obtained by the consumer-level depth camera Kinect,for obtaining the depth data of the scene and the corresponding RGB data.And then get the head direction and joint points of the human body through the Kinect SDK,and select the proper points corresponding to static human body model.Combining with the hand joint points obtained from the RGB image by the deep learning method,computing the whole body skeleton of the body.The RGB-image based human body segmentation algorithm is also applied obtain complete low noise human body point cloud data.Finally,the human body model is computed by fitting the human body fitting objective function with the pose keypoints and optimizing the details with point cloud data.Accurately estimating the human body shape in any posture under any clothes can be fulfilled by the proposed method.Finally,the accuracy of human joint data plays an important role in human body estimation method.Therefore,we propose a 3D joint estimation method by using both depth data method and RGB image method.The depth method uses the Kinect SDK's joint point prediction,however it may obtain specific joints that are not accurately captured,so that we can adjust the wrong joint points which can be corrected by the RGB image method.We use the recent neural network human 3D joint prediction method to update the incorrect joint points for an accurate 3D keypoint predicction.Experimental results show the efficiency of the proposed method.