Research On 3D Human Body Reconstruction Based On Single Frame Image And SMPL

The rapid advancement of computer hardware technology and deep learning has propelled the rapid growth of 3D human body reconstruction.Notably,scholars have extensively focused on the task of reconstructing a 3D human body from a single frame of RGB images.Although extant methods for 3D human body reconstruction based on single-frame RGB images have made significant strides,issues such as inadequate precision and lack of detail still exist.To tackle the above issues,this study centers on the subject of reconstructing a 3D human body from a single frame of RGB images and employs deep learning technology and the parametric human body model SMPL to accomplish 3D human body reconstruction.Specifically,the research objectives are twofold:(1)This article presents a novel approach to enhance the accuracy of 3D human body modeling using a multi-input method.The proposed method leverages intermediate representations,such as 2D joint points,to guide the output.Specifically,the original RGB image,2D joint point map,and 2D segmentation map are employed to perform 3D human body reconstruction.The methodology involves generating 2D joint and segmentation maps from the RGB image,followed by feature extraction using three encoding networks.The features are then concatenated and used as input to a regressor to obtain SMPL human body parameters.The final 3D human body model is generated using these parameters.The experimental results demonstrate that the proposed multi-input method significantly reduces the reconstruction error and enhances the reconstruction quality compared to the 3D human body reconstruction that solely relies on a single-frame RGB image.(2)To achieve a more accurate 3D human body reconstruction,this study proposes the use of the ResNeXt-CBAM encoding network,a more effective encoding network for 3D human body modeling.Since deep learning is employed to reconstruct the 3D human body,a powerful encoding network is crucial.The ResNeXt-CBAM network exhibits strong adaptability,small scale,strong scalability,and expressive ability,making it suitable for the 3D human body reconstruction task.To evaluate the efficacy of the proposed ResNeXt-CBAM encoding network,this study conducts several experiments on 3D human body reconstruction using different methods and compares its performance against other networks.The experimental results demonstrate that the ResNeXt-CBAM encoding network significantly enhances the performance of human body reconstruction and reduces the reconstruction error of 3D human body models.