3D Human Reconstruction Based On SMPL Model

In recent years,the iterative upgrade of hardware such as graphics processing units(GPUs)has accelerated the development of artificial intelligence technology and deep learning.3D human reconstruction based on deep learning has become a research hotspot in the field of computer vision,and is widely used in the fields of motion analysis and intelligent security,with great research value and practical value.3D pose estimation has made great progress in the field of computer vision.This paper takes 3D pose estimation as the basic task combined with Skinned Multi-Person Linear Model(SMPL)to achieve 3D human reconstruction,and at the same time studies how to improve the performance of 3D pose estimation algorithm to further help human reconstruction tasks.The main research work is as follows:1.Parametric human model provide rich prior knowledge to make 3D reconstruction tasks simple and effective,and this paper uses SMPL models to achieve human reconstruction.Aiming at the problem that the SMPL model parameters are highly nonlinear and the accuracy is low from the image,this paper proposes a 3D human reconstruction model based on joint features and graph convolutional network.The 3D pose estimation module is added to the model,and the joint features are extracted by combining the position of human joint points combined with bilinear interpolation to provide pixel-level positioning information for model prediction.In addition,the graph convolutional network is introduced to capture the implicit spatial information between features to improve the prediction accuracy of three-dimensional rotation.By narrowing the gap between the 3D joint point estimation and the joint point of the SMPL model,the rationality and accuracy of the human reconstruction results are guaranteed.The model was tested on the public datasets and the results show the validity of the model.2.In view of the lack of 3D data and the poor generalization of 3D pose estimators,this paper proposes a 3D pose estimation method based on pose augmentation to improve the accuracy of 3D pose estimation model by improving the diversity of 2D-3D pose pairs.The model consists of a generator,a discriminator and a 3D pose estimator,the generator and discriminator provide reasonable data for the training of the 3D pose estimator,and the 3D pose estimator feedback the error to the generator,prompting the generation of more difficult poses.The generator is composed of bone length and bone rotation generator,and the discriminator uses the local kinematic chain space as input to improve the diversity of postures while ensuring the rationality of the generated postures.The results of experiments on different 3D pose estimators show that the proposed method can improve the accuracy of different 3D pose estimators with good scalability.Finally,it is verified on the dataset that the accuracy of the 3D human reconstruction can be improved by improving the accuracy of the 3D pose estimation algorithm.