Research On 3D Human Modeling And Its Application Based On SMPL Model

The technology of 3D human modeling has a broad application prospect in fields such as development of games,animation production and virtual fitting.Due to the reasons like complicated and changeable body constitution and more exterior non-rigid body motion,the high-precision 3D human reconstruction technology mainly depends on the complicated and expensive multi-perspective camera system or laser scanner for data acquisition and processing currently.With the presentation of SMPL model,using the human prior constraints provided by model can transform the complicated problem of 3D reconstruction into a series of parameter optimization problems,so how to use the constraints provided by SMPL model to make the effective,convenient and fast 3D digitalization reconstruction has become an issue being discussed by scholars.In the above context,the task mainly studies how to use the incomplete information gathered with the single sensor device to make the optimization solution to the model parameter centering on the subject of 3D human modeling based on SMPL model,as the specific jobs can be summarized as follows:(1)Human posture reconstruction based on the single RGB image.To solve the current problems such as that an error occurs easily in the posture reconstruction algorithm based on the SMPL model,and the incorrect model parameter estimation,the article combines the posture reconstruction algorithm based on optimization with the one based on deep learning:Firstly to extract the model parameter from the image as the original value with the deep convolutional neural network,and then use the constraints such as human silhouette image,2D articulation point and human height to make the further optimization solution of the model parameter,and obtain the more precise 3D postures and types of human body.To verify the effectiveness of the proposed methods,the algorithms of the same kind are evaluated in the truthful data and public data sets,as the experiment shows that the proposed methods can effectively combine the advantages of both optimization algorithm and deep learning algorithm to improve the precision of 3D human posture reconstruction algorithm.The experiment also shows the application effects of body parts analysis for the proposed methods to support the related applied research.(2)Human posture reconstruction integrated with the RGBD information.To solve the problem of non-uniqueness of posture solution caused by the lack of depth information of 2D image by integrating the depth information based on the human posture reconstruction on account of the single RGB image:Firstly to gather the single-frame human color image and depth image with the Kinect camera and generate the point cloud data,and then make the operations to the point cloud such as background segmentation and denoising to obtain the high-quality human unilateral point cloud;To use the 3D human posture reconstruction algorithm based on the RGB image to obtain the initial model and make the coarse registration of it with the point cloud;Finally to confirm the point-to-point correspondence between the initial model and human point cloud with the K-nearest neighbor algorithm,and make the optimization of model parameter with the distance between the point cloud and model to precisely estimate the postures and types of human body under the complicated postures.Methods in the article are verified to have the features such as high precision and simple experimental environment,and the precise 3D postures and types of human body can be obtained through the qualitative and quantitative analysis of truthful somatic data collected from the experiment.And finally to demonstrate the application effects of virtual fit combined with the simulation software of garment cloth to the proposed algorithms.(3)Human dynamic reconstruction based on the RGB video sequence.As an error occurs easily in the human posture reconstruction algorithm based on the single RGB image during the self-occlusion of human body,problems such as incoherence of movement occurs easily while handling the video.A 3D human dynamic reconstruction method based on the RGB video sequence is adopted to solve the above problems and make the SMPL model able to contain much more texture details:Firstly to make the human posture reconstruction to each frame of RGB video sequence,and then make the error correction to the posture reconstruction results of all frames in the video with the frame-to-frame consistency of video sequence to make the component motional sequence much more precise and fluent;And to transform the cone of human silhouette in each frame into the standard T-type posture with the deformation principle of SMPL model,and finally make the optimization solution of vertex offset of SMPL model with the constraints of visual hull consisted by cones of various outlines to obtain a human model containing the texture information.Through the experiment with the human data gathered from the truthful environment,the result shows that the proposed methods can make the 3D reconstruction for the dynamic human body with any simple postures.And due to that the animation drive of generated human model can be animated by parameter change,the experiment also shows the application demonstration of action imitation in combination with the proposed 3D human posture reconstruction algorithm.