Monocular human body three-dimensional attitude estimation method based on data enhancement architecture

Abstract

The invention belongs to the technical field of computer graphics, and relates to a three-dimensional human body attitude estimation method, which can accurately regress to obtain a three-dimensional posture only through a two-dimensional posture obtained by a single image, gets rid of the technical constraint that the accurate three-dimensional posture can be obtained by depending on high-cost hardware, and improves the accuracy of human body attitude estimation. According to the method, the cost of human body three-dimensional posture-dependent applications such as human-computer interaction, augmented reality and virtual reality is greatly reduced. Meanwhile, the three-dimensional posture capturing precision of unusual actions is greatly expanded, more complex actions can appear in the applications such as human-computer interaction, and the interestingness and practicability of the applications are greatly improved; the method is scientific and reliable in principle, data diversity is expanded on the basis of an existing data set, the model generalization ability is improved, vivid and natural three-dimensional human body postures are obtained by relying on image data collected by a monocular camera in a richer real scene, the precision of three-dimensional attitude estimation of unusual actions can be remarkably improved, and the method is suitable for popularization and application. The method can be applied to more diversified scenes.

Description

Technical field: [0001] The invention belongs to the technical field of computer graphics, and relates to a method for estimating a three-dimensional human body posture, in particular to a monocular three-dimensional human body posture estimation method based on a data enhancement framework. Background technique: [0002] With the continuous development of human body pose estimation research and application, 2D human body pose estimation based on monocular images has achieved remarkable results. In the 3D world, 3D human body pose can provide more realistic and three-dimensional sensory effects. Attitude needs to be obtained with the help of professional depth cameras or body-worn sensors, and the application threshold is too high. [0003] With the help of the two-dimensional image collected by the monocular camera, the three-dimensional pose is generated through neural network regression, which can greatly improve the convenience of application, expand more application pos...

AI Technical Summary

Problems solved by technology

However, due to the lack of depth information in 2D images, there is inherent depth ambiguity in mapping from 2D images to 3D human poses, and a 2D image can correspond to multiple 3D human poses, so the mapping also has uncertainty

[0004] In the prior art, research on 3D human pose estimation based on a monocular camera needs supervised or weakly supervised training through a data set containing accurate 3D poses and corresponding 2D images as input. However, in order to ensure the accuracy of the data, the data set It needs to be collected in a laboratory equipped with professional cameras and sensors, and it is completed by a dozen people simulating multiple specific scenarios. The neural network model trained based on these data sets will show generalization when faced with real outdoor application scenarios Insufficient problems, and at the same time, the effect is not ideal for some uncommon actions (falling and flipping, etc.)

[0005] Therefore, it is necessary to develop and design a 3D human pose estimation method based on monocular vision that can effectively solve the problem of insufficient data diversity and poor model generalization ability.

Images