A Human 3D Pose Estimation Method Combining Densely Connected Attention Pyramid Residual Networks and Isometric Constraints

Abstract

The invention discloses a human body 3D posture estimation method combined with a densely connected attention pyramid residual network and isometric constraints. The method is decomposed into two parts: a discriminative human body 2D posture estimation and a generative human body 3D posture estimation. First construct a human body 2D pose estimation model. The human body 2D pose estimation model includes an attention pyramid residual block and an hourglass sub-network composed of several attention pyramid residual blocks. The attention pyramid residual block is used for multi-scale image feature extraction, and hourglass The sub-network is used to generate the heat map of human joint points; in order to solve the problem of underutilization of environmental context information, the attention mechanism and multi-scale analysis are combined to capture the environmental context features; in order to solve the problem of gradient disappearance / gradient explosion, the densely connected network combines the above attention The force mechanism improves the recognition of feature maps. Then construct the loss function, introduce the isometric restriction term, and fit the human body 3D pose by minimizing the loss function. The method of the invention has obvious advantages in the task of estimating the 3D posture of the human body.

Description

technical field [0001] The invention belongs to the technical field of human body posture estimation, and in particular relates to a human body 3D posture estimation method combined with a densely connected attention pyramid residual network and isometric constraints. Background technique [0002] Human 3D pose estimation recovers the 3D positions of human joint points in a given image or video. This work is the basis for many important applications, such as video surveillance, behavior recognition, human interaction, virtual reality, game animation, medical care, and more. [0003] The current human pose estimation methods can be roughly divided into the following categories: 1) Regression iterative method, initialization to get the initial pose prediction, and then iterative estimation to improve the prediction accuracy; 2) Method based on structured learning, using Markov random field mining Human body structure information to obtain the mutual relationship of human body...

AI Technical Summary

Problems solved by technology

[0004] At present, the research on human 3D pose estimation based on deep convolutional neural network has achieved certain results, but there are some bottlenecks in performance: 1) This problem is essentially a pathological problem; 2) The mapping from image feature space to 3D pose space is wrong. Linear multimodal; 3) Deeper networks are easy to learn this nonlinear mapping relationship, but deeper networks are prone to gradient disappearance or gradient explosion problems

Images