Three-dimensional human body posture estimation method and system based on deep learning

Abstract

The invention discloses a three-dimensional human body posture estimation method and system based on deep learning, and the system comprises an image obtaining module which obtains an image, and a two-dimensional joint extraction module which carries out the two-dimensional joint extraction of the obtained image, and obtains a two-dimensional joint; using a joint point conversion module for carrying out joint point conversion on the two-dimensional joints obtained by the two-dimensional joint extraction module; and using a three-dimensional joint extraction module and a three-dimensional joint pre-training module for carrying out joint deep learning training on the two-dimensional joints subjected to joint point transformation by the joint point transformation module, and extracting three-dimensional human body postures; automatically learning transformation parameters, the invention is more suitable for the two-dimensional attitude transformation process, and by limiting the transformation process, direct adaptive transformation can be carried out on coordinate points of the two-dimensional attitude, and the problem that errors are too large in the deep learning process is solved.

Description

technical field [0001] The present invention relates to the technical field of three-dimensional human body posture, in particular to a method and system for estimating three-dimensional human body posture based on deep learning. Background technique [0002] Three-dimensional human body posture is a very important part of the existing computer vision. Generally speaking, human body posture is divided into three important researches: extraction from image to 2D posture, extraction from image to 3D posture, and extraction from 2D posture to 3D posture field. In the process of 3D human body extraction, the extraction accuracy directly from images to 3D poses is poor, so the research of this application is based on the extraction of images to 2D poses and the extraction of 2D poses to 3D poses. [0003] In the current computer vision, data enhancement is almost always applied at the image level. For example, in the paper "Adversarial Semantic Data Augmentation for Human Pose E...

AI Technical Summary

Problems solved by technology

However, it is obvious that this research is a direct mathematical operation, which is equivalent to performing a mathematical transformation on the original data set. Through the formula given in the article, it is a fixed transformation of the original two-dimensional human body posture, so as to obtain the transformed Human body posture cannot be given to adapt to the network parameters to learn two-dimensional human body posture, and it is even more impossible to give the corresponding three-dimensional human body posture as a supervisory signal for the new network to learn. It is pointed out in the article that when the amount of data meets the conditions, The effect in this paper cannot reach the effect of "3D human pose estimation in video with temporal convolutions and semi-supervised training" in 2019. It is difficult to give an adaptive data enhancement scheme for data enhancement in this direction.

Images