Three-dimensional human body posture and shape reconstruction method capable of shielding perception

Abstract

According to a three-dimensional human body posture and shape reconstruction method oriented to shielded human body scenes, a human body model can be reconstructed from a shielded human body scene, so that a computer can understand and analyze the human body model conveniently, the human body model is not influenced by a picture scene, and an accurate human body model is reconstructed. The method provided by the invention comprises the following steps: step 1, generating a human body contour map of mutually shielded human bodies; step 2, coding human body joint points by using the human body contour map generated in the step 1 to serve as input data; step 3, obtaining human body model parameters by using a regression network composed of a ResNet encoder and a regression device; step 4, rendering and generating a pre-estimated human body contour and human body joint points by using pre-estimated human body parameters; step 5, comparing the pre-estimated human body contour and human body joint points with real labels, carrying out reverse transmission, and calculating loss functions; step 6, setting an accuracy rate, and repeating the steps 2 to 5 until the set accuracy rate is reached.

Description

technical field [0001] The invention relates to the field of computer vision, in particular to a method for reconstructing three-dimensional human body posture and shape from a single picture. Background technique [0002] With the enhancement of computer computing power, deep learning has achieved great success, and it has also greatly promoted the computer's understanding ability. Computers need the support of human pose estimation in intelligent creation, intelligent medical care, automatic driving and game modeling. At present, great success has been achieved in human body 2D posture detection, which can accurately determine human body joint points in different carriers, such as pictures and videos. However, in 3D pose and shape reconstruction, due to the lack of depth information in the picture, the prediction of 3D pose lacks certain accuracy. In addition, in the real world, the existence of various complex scenes (such as human occlusion, object occlusion, etc.) wil...

AI Technical Summary

Problems solved by technology

In addition, the measured objects in various data sets have a single body type, and each data set may only contain a few individuals, which makes the diversity of the human body lacking. The model trained in this way may lack sensitivity to the human body, making it difficult to understand the human body. well reconstructed

Images