Research On 3D Reconstruction Of Human Body Based On 2D Images In Natural Scenes

3D reconstruction of the human body based on 2D images aims to estimate the 3D model of a person from 2D images.It has broad application prospects in fields such as human-computer interaction and film production,and is a valuable research direction and hotspot in academic research.The current commonly used method for 3D reconstruction of the human body based on 2D images is to use deep learning or optimization learning to extract parameters such as body shape and posture from 2D images.These parameters are then used to drive parametric human models to obtain a 3D human model with a consistent posture and body shape with the person in the image.In recent years,with the emergence of large-scale 3D human pose datasets and the continuous development of deep learning techniques,the performance of algorithms has been significantly improved.However,there are still problems with the current methods:1)The deep learning algorithm’s generalization ability is limited due to the difficulty of covering the complex and diverse natural scenes with training data.It’s challenging to extract robust human body features from natural scene images,and the parameter regression method that depends on global features is also sensitive to partial occlusion.2)It is difficult to obtain depth information from a monocular image,and the lack of depth supervision can cause depth blurring issues.This makes it difficult to effectively constrain pose parameters in optimization algorithms and affects the reconstruction quality.To address these problems,the following work is carried out in this thesis:(1)To address the problem of the limited generalization ability of deep learning algorithms,this thesis designs a human body 3D reconstruction method based on feature enhancement and local features.By using easily detectable 2D poses to enhance image features,the network is guided to accurately extract features of the target human body from the image,thereby improving the algorithm’s ability to extract human body features in natural scene images.Additionally,by decomposing human body features into the local features of multiple joints,the impact of occlusion is restricted within a local range.A parameter regression network based on local joint features was designed to model the interdependence of local features,thereby improving the accuracy of parameter prediction.(2)To address the depth blurring issue in optimization algorithms,this thesis designs a human body 3D reconstruction method based on motion priors.The motion prior models the regularity of human body motion in the video and is pre-trained on the human motion dataset to generate a future human posture sequence based on past posture sequences.The motion prior can provide pose parameters that conform to the regularity of human body motion to the optimization algorithm.By using the temporal consistency of the posture sequence,the pose parameters are constrained within a reasonable range to alleviate the pose error caused by depth blurring and improve the reconstruction quality.