An enhanced real-time human pose estimation method based on modified YOLOv8 framework.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning-based human pose estimation (HPE) aims to predict human body posture in images/videos.
• Real-time HPE accuracy is limited by partial occlusion and restricted model receptive fields.
• Existing models struggle with feature loss and receptive field constraints, impacting pose accuracy.

Purpose of the Study:

• To propose a novel real-time human pose estimation model, CCAM-Person, based on the YOLOv8 framework.
• To enhance the accuracy of HPE by addressing feature loss and receptive field limitations.
• To improve robustness against occlusions and background noise for more precise keypoint regression.

Main Methods:

• Modified the backbone and neck of the YOLOv8x-pose model to mitigate feature loss and receptive field issues.
• Introduced the Context Coordinate Attention Module (CCAM) to enhance focus on salient features.
• Evaluated the CCAM-Person model on the MS COCO 2017 and CrowdPose datasets.

Main Results:

• The CCAM-Person model demonstrated competitive performance on multiple metrics across both datasets.
• Achieved a 2.8% and 3.5% improvement in average precision on MS COCO 2017 and CrowdPose, respectively, compared to YOLOv8x-pose.
• The CCAM module effectively reduced background noise and improved keypoint regression accuracy, especially with limb occlusion.

Conclusions:

• The proposed CCAM-Person model significantly improves real-time human pose estimation accuracy.
• The enhancements to the YOLOv8x-pose framework, particularly the CCAM module, effectively address key challenges in HPE.
• This approach offers a more robust and accurate solution for real-world HPE applications.