A feature interaction fusion method and system for 3D human pose estimation

By employing a global-local feature interaction fusion method, and utilizing a multi-level attention mechanism to stack encoders and optimize loss functions, the problem of insufficient inter-frame information in 3D human pose estimation is solved, thus improving detection accuracy. This method is suitable for applications such as human-computer interaction and motion capture.

CN117115915BActive Publication Date: 2025-12-12HUNAN UNIV
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202311138143.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-09-05
Publication Date
2025-12-12
Estimated Expiration
2043-09-05

AI Technical Summary

Technical Problem

Existing 3D human pose estimation techniques do not make sufficient use of inter-frame information, resulting in large differences in detection results, making it difficult to improve the accuracy of the algorithm and limiting its application in high-precision scenarios.

Method used

A feature interaction fusion method is adopted, which obtains global and local features through a multi-level attention mechanism stacked encoder, and combines loss function to optimize the model to achieve effective fusion of global and local features.

Benefits of technology

It improves the accuracy of 3D human pose estimation, can effectively utilize inter-frame information, and is suitable for applications such as human-computer interaction and motion capture.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117115915B_ABST
    Figure CN117115915B_ABST
Patent Text Reader

Abstract

The application relates to the technical field of 3D human posture recognition, and discloses a feature interaction fusion method for 3D human posture estimation, which comprises the following steps: acquiring a 2D posture sequence corresponding to a video to be predicted, improving the data dimension of each frame of data in the 2D posture sequence, obtaining X in ; inputting the acquired X in into different encoders based on a multi-level attention mechanism stack in the form of feature blocks, respectively acquiring global layer features and local layer features; fusing the acquired global layer features and local layer features to obtain fusion features; constructing a 3D human posture sequence according to the fusion features and coordinate regression, acquiring the 3D human posture of a human body in the video to be predicted according to the 3D human posture sequence, and simultaneously focusing on the local layer and the global layer through the design of a double-flow design and a multi-level attention, so that the precision of 3D human posture detection is improved.
Need to check novelty before this filing date? Find Prior Art

Citation Information

Patent Citations

  • Self-adaptive multi-view and time sequence feature fusion single-person three-dimensional attitude estimation method

    CN114511629A

  • Multi-view feature fusion method and system for 3D human body posture estimation

    CN114758205A