Prediction of lower limb kinematics from vision-based system using deep learning approaches

Area of Science:

• Biomechanics
• Computer Vision
• Machine Learning

Background:

• Joint angular velocity is crucial for assessing injury risk, rehabilitation progress, and athletic performance.
• Wearable sensors are commonly used for lower limb kinematics but are inconvenient for daily use.

Purpose of the Study:

• To develop and validate a vision-based system for estimating lower limb joint angular velocity without physical sensors.
• To evaluate the accuracy of deep learning models in capturing joint kinematics during daily life exercises.

Main Methods:

• A deep convolution neural network combined with bi-directional long-short term memory and gated recurrent unit networks was employed.
• The system estimated joint angular velocities, with results compared against an optical motion capture system.
• Normalized correlation coefficient and mean absolute error were used as evaluation metrics.

Main Results:

• The vision-based system achieved high correlation coefficients (0.93 for squat, 0.92 for treadmill walking).
• The knee joint exhibited the highest estimation accuracy (0.96 for squat and treadmill walking).
• Accurate estimations were maintained across different camera views and positions.

Conclusions:

• A sensor-free, vision-based system can effectively monitor lower limb kinematics during home workouts.
• The proposed deep learning models provide a practical and accurate solution for healthcare and rehabilitation applications.
• This technology enhances the convenience of kinematic monitoring for physical activity analysis.