Lightweight human pose estimation and abnormal behavior recognition method for unmanned aerial vehicle aerial scene

By employing multi-scale target detection and semantically aware RoI feature extraction, keypoint adaptive multi-scale fusion, and lightweight manifold dynamic graph network, the problem of perspective change and scale difference in human pose estimation and abnormal behavior recognition in UAV aerial photography scenarios is solved, achieving high-precision real-time detection and recognition.

CN122416307APending Publication Date: 2026-07-17SHENYANG AEROSPACE UNIVERSITY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202610533691.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-04-22
Publication Date
2026-07-17

AI Technical Summary

Technical Problem

Human pose estimation and abnormal behavior recognition in drone aerial photography scenarios face challenges such as insufficient feature extraction due to changes in perspective, lack of targeted multi-scale feature fusion, and difficulty in balancing computational efficiency and accuracy. In particular, it is difficult to achieve real-time high-precision detection on drone platforms with limited computing resources.

Method used

A multi-scale target detection network is adopted, combined with a semantic-aware RoI feature extraction module and a key point adaptive multi-scale feature fusion module. A key point heatmap is generated through a lightweight high-resolution network, and the key point coordinates are converted into a manifold space representation. Spatiotemporal modeling is performed in combination with a lightweight manifold dynamic graph network to achieve abnormal behavior recognition.

Benefits of technology

It significantly improves attitude estimation accuracy and abnormal behavior recognition accuracy. The model is lightweight and has strong real-time performance. It is adaptable to the top-down view of UAVs and small-scale targets, meets the requirements of UAV airborne deployment, and has good robustness and engineering deployment value.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122416307A_ABST
    Figure CN122416307A_ABST
Patent Text Reader

Abstract

This invention discloses a lightweight human pose estimation and abnormal behavior recognition method for UAV aerial photography scenarios. First, a YOLOv11 multi-scale detection network is used for human target detection to obtain bounding boxes and multi-scale feature maps. Then, a semantically aware RoI feature extraction module is used to extract semantically consistent region features using dynamic sampling and learnable semantic masks. Next, a keypoint adaptive multi-scale fusion module is employed to dynamically adjust the scale fusion weights based on prior knowledge of human anatomy. The fused features are then input into a lightweight high-resolution network equipped with a one-dimensional spatial attention mechanism for heatmap regression and keypoint coordinate decoding. Finally, the continuous frame keypoint sequence is input into a lightweight manifold dynamics graph network, combining quaternion skeleton representation, single-step Eulerian dynamics, and sparse grouped graph convolution for abnormal behavior recognition. This invention can improve the accuracy of pose estimation from the UAV's perspective and achieve real-time detection of abnormal behaviors such as falls and fights.
Need to check novelty before this filing date? Find Prior Art