Three-dimensional human body model reconstruction method, storage equipment and control equipment

Abstract

The invention relates to the technical field of human body model reconstruction, in particular to a three-dimensional human body model reconstruction method, storage equipment and control equipment, and aims to solve the problems that a three-dimensional model reconstruction result is not aligned with a two-dimensional image and the like. The reconstruction method comprises the steps that according to a human body image, adopting a pre-trained full convolutional network module to acquire a global UVI image and a local UVI image of a human body part; estimating camera parameters and shape parameters of the human body model by using a first neural network based on the global UVI graph; based on the local UVI graph, utilizing a second neural network to extract rotation features of each articulation point of the human body; based on the rotation characteristics of each articulation point of the human body, improving the rotation characteristics by utilizing a characteristic improvement strategy based on position assistance to obtain improved rotation characteristics; and estimating attitude parameters of the human body model by using a third neural network according to the improved rotation characteristics. According to the method, the human body model can be reconstructed more accurately and efficiently, and the robustness of attitude estimation is improved.

Description

technical field [0001] The invention relates to the technical field of human body model reconstruction, in particular to a three-dimensional human body model reconstruction method, a storage device, and a control device. Background technique [0002] 3D human body model reconstruction is one of the important tasks of 3D computer vision, which aims to quickly and accurately reconstruct a 3D human body parametric model from human body images, including the shape parameters and posture parameters of the model. [0003] In the existing 3D human body model reconstruction technology, the traditional model-based fitting method usually deforms the 3D human body model iteratively, so that the reprojected model matches the 2D image information such as the position of joint points or the contour of the human body. Time consuming and sensitive to initial values ​​of model parameters. The emerging learning-based methods use neural networks to directly extract features from images and es...

AI Technical Summary

Problems solved by technology

[0003] In the existing 3D human body model reconstruction technology, the traditional model-based fitting method usually deforms the 3D human body model iteratively, so that the reprojected model matches the 2D image information such as the position of joint points or the contour of the human body. Time consuming and sensitive to initial values ​​of model parameters

The emerging learning-based methods use neural networks to directly extract features from images and estimate model parameters. These methods have improved the accuracy of model shape and pose estimation to a certain extent, but there are still differences between the reconstruction results of 3D models and 2D images. Alignment etc.

There are two main reasons for this: (1) there is a highly nonlinear mapping relationship between the image and the model parameters; (2) the pose of the 3D human body model is usually expressed by the relative rotation of the joint points, which leads to the reprojection of the model reconstruction results in the Positional deviations are prone to exist on 2D images

Images