Research on Human Posture Estimation Algorithm Based on YOLO-Pose.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Traditional human pose recognition methods struggle with challenges like occlusion, dense targets, and complex backgrounds.
• Existing algorithms often have limited application scenarios and poor accuracy.

Purpose of the Study:

• To propose an improved YOLO-Pose algorithm for robust and accurate human pose estimation.
• To enhance performance for practical applications, including deployment on unmanned aerial vehicles (UAVs).

Main Methods:

• Integrated lightweight GhostNet modules in the Backbone for reduced parameters and computation.
• Incorporated ACmix attention mechanism in the Neck for faster detection and localization.
• Optimized key points in the Head using coordinate attention and improved loss/confidence functions for robustness.

Main Results:

• Achieved a 95.58% improvement in mAP50 and 69.54% in mAP50-95 compared to the original model.
• Reduced model parameters by 14.6 M.
• Increased detection speed to 19.9 ms per image, a 30% and 39.5% optimization.

Conclusions:

• The enhanced YOLO-Pose algorithm significantly improves human pose estimation accuracy and efficiency.
• The model demonstrates superior performance and robustness compared to existing methods like Faster R-CNN, SSD, YOLOv4, and YOLOv7.