A human pose estimation network based on YOLOv8 framework with efficient multi-scale receptive field and expanded...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Traditional multi-person pose estimation struggles with occlusions and body part overlaps.
• Deep learning-based methods offer advanced solutions for human pose estimation in images and videos.

Purpose of the Study:

• To introduce EE-YOLOv8, a novel human pose estimation network designed to overcome limitations of existing methods.
• To enhance feature representation, optimize multi-scale data integration, and improve detection accuracy.

Main Methods:

• Developed EE-YOLOv8, integrating Efficient Multi-scale Receptive Field (EMRF) and Expanded Feature Pyramid Network (EFPN) into the YOLOv8 framework.
• Employed Wise-IoU to replace traditional Intersection over Union (IoU) for more precise bounding box overlap measurement.
• Evaluated performance on the MS COCO 2017 dataset.

Main Results:

• EE-YOLOv8 achieved an AP of 89.0% at an IoU threshold of 0.5, a 3.3% improvement over YOLOv8-Pose.
• Achieved an AP of 65.6% over the IoU range of 0.5-0.95, a 5.8% improvement.
• Demonstrated superior accuracy with the lowest parameter count and computational complexity.

Conclusions:

• EE-YOLOv8 significantly outperforms existing mainstream human pose estimation methods.
• The proposed network offers a highly accurate and computationally efficient solution for complex pose estimation tasks.
• EE-YOLOv8 shows strong competitiveness in the field of deep learning-based human pose estimation.