A Method for Human Pose Estimation and Joint Angle Computation Through Deep Learning

Area of Science:

• Computer Vision
• Medical Imaging
• Biomechanical Analysis

Background:

• Human pose estimation is vital for healthcare, sports, and remote monitoring.
• Accurate joint angle computation is essential for analyzing body posture and alignment in physiotherapy.

Purpose of the Study:

• To develop a deep learning method for automatic human pose estimation and joint angle computation.
• To tailor the method for physiotherapy and telemedicine applications, enabling motion analysis.
• To evaluate the method's effectiveness in real-world use cases for exercise and posture assessment.

Main Methods:

• A deep learning approach utilizing a customized 25-anatomical keypoint skeleton.
• Training on a large dataset of over 150,000 annotated and augmented images from open-source datasets.
• Implementation for both keypoint localization and object detection.

Main Results:

• Achieved a mean Average Precision (mAP@50) of 0.58 for keypoint localization.
• Achieved a mAP@50 of 0.98 for object detection.
• Demonstrated practical use cases in evaluating exercise correctness and identifying postural deviations.

Conclusions:

• The proposed method offers a promising approach for automated motion analysis in digital health.
• It has potential impact on rehabilitation support and remote patient care through accurate pose and angle computation.
• The method effectively supports physiotherapy by analyzing exercise correctness and postural deviations.