Human posture estimation model training method and human posture estimation method

By performing multi-frame image processing and multi-task learning on the human pose estimation model, the problems of insufficient temporal continuity and spatial positioning accuracy in traditional methods are solved, and accurate estimation and tracking of human pose key points are achieved, improving detection accuracy and robustness.

CN122199664APending Publication Date: 2026-06-12DONGFENG HONDA ENGINE CO LTD +1
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
DONGFENG HONDA ENGINE CO LTD
Filing Date
2026-02-10
Publication Date
2026-06-12

AI Technical Summary

⚠Technical Problem

Traditional human pose estimation methods suffer from problems such as neglecting temporal continuity and insufficient spatial positioning accuracy in single-frame images and video sequences, resulting in inaccurate human pose key point detection and estimation.

⚗Method used

By performing multi-frame image processing on sample videos containing human poses, a heatmap of predicted key points is generated. Initial estimated coordinates are obtained and refined temporally and spatially. The human pose estimation model is optimized by combining a multi-task learning strategy, achieving temporal-spatial co-optimization.

🎯Benefits of technology

It significantly improves the accuracy of human pose key point detection, reduces temporal jitter between adjacent frames, enhances the smoothness of motion trajectory and spatial positioning accuracy, and strengthens robustness in complex scenes.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122199664A_ABST
    Figure CN122199664A_ABST
Patent Text Reader

Abstract

The application relates to a human posture estimation model training method, a human posture estimation method and a human posture estimation model training device. The application comprises the following steps: inputting multiple target sample images in a sample video carrying a human posture into a human posture estimation model to be trained to generate a predicted key point heat map of each target sample image; obtaining initial estimated coordinates of each key point in each target sample image from each predicted key point heat map, and obtaining predicted time sequence refined coordinates according to the initial estimated coordinates; obtaining predicted target refined coordinates according to the predicted key point heat map; obtaining a human posture estimation loss according to the predicted key point heat map, the predicted time sequence refined coordinates and the predicted target refined coordinates; and training the human posture estimation model to be trained to obtain a trained human posture estimation model. The application can improve the accuracy of human posture key point detection estimation.
Need to check novelty before this filing date? Find Prior Art