KSL-POSE: A Real-Time 2D Human Pose Estimation Method Based on Modified YOLOv8-Pose Framework.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Real-time 2D human pose estimation faces challenges with occluded or overlapped individuals, reducing accuracy.
• Existing methods struggle with precise keypoint localization in complex scenarios.

Purpose of the Study:

• To develop an improved 2D human pose estimation method addressing accuracy limitations in challenging image conditions.
• To enhance feature extraction and small keypoint detection within the YOLO framework.

Main Methods:

• Proposed KSL-POSE method integrates Kolmogorov-Arnold Networks (KANs) for enhanced convolutional feature extraction.
• Incorporated cross-stage partial (CSP) approach and a small object enhance pyramid (SOEP) module for improved small target detection.
• Introduced a layered shared convolution with batch normalization detection head (LSCB) for efficient feature fusion and parameter utilization.

Main Results:

• KSL-POSE achieved a 1.5% increase in average detection accuracy compared to YOLOv8l-POSE on the MS COCO 2017 dataset.
• Demonstrated competitive performance on the CrowdPOSE dataset, indicating strong generalization ability.
• The method effectively handles occlusions and overlaps, improving overall pose estimation accuracy.

Conclusions:

• KSL-POSE offers a significant advancement in 2D human pose estimation, particularly for complex scenes.
• The integration of KANs, CSP, SOEP, and LSCB modules enhances model performance and efficiency.
• The proposed method shows promise for real-world applications requiring robust human pose recognition.