2D human body pose generation method and device for strong interaction human body motion

Abstract

The invention discloses a 2D human body pose generation method for strong interaction human body motion, and the method mainly comprises the steps: collecting and building a strong interaction human body motion data set through a built multi-view camera array; performing fine tuning on a pre-training model of the 2D human body pose estimation network HRNet based on the established data set so as to improve the detection capability of the pre-training model on strong interactive motion; the model obtained through fine adjustment is used for detecting images of different visual angles, and initial features of the images are obtained; calculating corresponding polar line positions of the initial features of each view angle in other view angles according to the epipolar geometry principle; k features are sampled on the corresponding epipolar lines respectively, and feature similarity calculation and feature fusion are carried out on the K features and the source view angle to make up for the deficiency of the initial features of the source view angle; and adopting a greedy matching algorithm and using a skeleton length as a threshold to constrain connection of joint points, and finally generating a 2D human skeleton. According to the invention, the detection accuracy of the 2D human body pose under strong interaction human body motion can be effectively improved.

Description

technical field [0001] The invention belongs to the technical field of human body pose detection, in particular to a 2D human body pose generation method for strong interactive human motion. Background technique [0002] 2D human pose generation is a basic problem in the field of computer vision and graphics. It mainly uses ordinary cameras or depth cameras to obtain image information to predict the position of each joint point of the human skeleton, and connect to generate a 2D human skeleton. The foundation of technologies such as motion recognition, virtual fitting and human-computer interaction, but also has many application values ​​in the fields of medical care, competition, animation and smart cities. [0003] At present, 2D human pose generation methods can be divided into single-person 2D human pose estimation and multi-person 2D human pose estimation according to the number of objects. Due to the simplicity of the detection scene for single-person 2D pose estimati...

AI Technical Summary

Problems solved by technology

However, this kind of commercial system must be able to work in a specific environment, and the scope of application is strictly limited

In addition, there is a set of unmarked system methods in the research of 2D human pose estimation, in which a single camera is used to collect video, but the images collected in this way contain less information, especially when the characters perform strong interactive movements and there are serious occlusions has a serious impact on the accuracy of the test results

Images