Three-dimensional human body posture reconstruction method based on L1/2 regularization

Abstract

The invention discloses a three-dimensional human body posture reconstruction method based on L1 / 2 regularization. The method comprises the following steps: forming a database through three-dimensional coordinates of all nodes in a plurality of three-dimensional human body graphics, and performing dictionary learning on the database by using an online learning method of matrix decomposition and sparse coding to obtain an over-complete dictionary; then constructing a shape space model by using the over-complete dictionary; then, performing convex relaxation processing on a non-convex optimization problem by using the properties of a spectral norm and the characteristics of L1 / 2 regularization to convert the non-convex optimization problem into a convex programming problem; then, convertingthe convex programming problem into an augmented Lagrangian solution expression; then, performing iterative solution on the augmented Lagrangian solution expression by using the ADMM algorithm; and finally, reconstructing a three-dimensional human body posture according to a solution value by using the shape space model and a 3D variable shape model. The three-dimensional human body posture reconstruction method has the advantages that the reconstruction effect is good and the sparsity is good.

Description

technical field [0001] The present invention relates to a three-dimensional reconstruction technology, in particular to a method based on L 1 / 2 A Regularized 3D Human Pose Reconstruction Method. Background technique [0002] Reconstructing the 3D shape of an object from a 2D image is one of the challenging problems in computer vision. In recent years, researchers' research focus has shifted from estimating the 3D shape of objects using edge boxes in images to further analyzing the 3D geometric information (such as shape, pose, etc.) of objects. Reasoning based on 3D geometric information can not only provide richer information for high-level vision tasks such as scene understanding, enhanced display, and human-computer interaction, but also can effectively improve object recognition performance. [0003] Reconstructing the 3D shape of an object from a single-view 2D image is an ill-conditioned problem in itself. In recent years, more and more researchers have used the eve...

AI Technical Summary

Problems solved by technology

Although the method proposed by Zhou Xiaowei and others successfully solved the problem of sensitivity to the initial value and noise, the essence is to use L 1 The regularized convex relaxation method transforms the non-convex optimization problem into a convex programming problem, but its disadvantage is that the reconstruction effect and sparsity are not ideal

Images