Deep Learning Networks for View-independent Knee and Elbow Joint Angle Estimation

Area of Science:

• Biomedical Engineering
• Computer Vision
• Machine Learning

Background:

• Vision-based human joint angle estimation is vital for remote and continuous health monitoring.
• Current methods often rely on optical or depth cameras and human pose estimation models.
• There is a need for reliable and straightforward vision-based methods using standard RGB video.

Purpose of the Study:

• To propose a deep learning-based approach for estimating knee and elbow flexion/extension angles from RGB video.
• To evaluate the performance of different deep learning networks for this task.
• To demonstrate a reliable and straightforward method for joint angle estimation.

Main Methods:

• Utilized deep learning networks, specifically BiLSTM, to estimate joint angles from RGB video.
• Input consisted of nine subsequent video frames.
• Output was compared against an optical motion capture system for validation.
• Experiments involved fifteen healthy participants performing daily activities.

Main Results:

• The BiLSTM network demonstrated high accuracy in estimating both knee and elbow joint angles.
• Achieved a correlation of 0.955 for knee joint angles and 0.917 for elbow joint angles.
• The estimation performance was consistent across different camera view angles.

Conclusions:

• The proposed BiLSTM network offers a reliable and straightforward method for vision-based knee and elbow joint angle estimation from RGB video.
• This approach holds significant potential for remote and continuous health monitoring applications.
• The method's robustness to camera angles enhances its practical applicability.