Supervised learning for improving the accuracy of robot-mounted 3D camera applied to human gait analysis

Area of Science:

• Robotics and Biomechanics
• Computer Vision
• Machine Learning

Background:

• 3D cameras have historically shown low accuracy in human gait analysis.
• Mobile robot platforms offer potential for extended gait data collection.
• Supervised learning presents a method to improve 3D camera performance.

Purpose of the Study:

• To enhance the accuracy of robot-mounted 3D cameras for human gait analysis.
• To investigate the application of supervised learning for improving gait estimations.
• To compare two distinct supervised learning approaches for gait analysis.

Main Methods:

• Utilized an Orbbec Astra 3D camera mounted on a mobile robot.
• Collected gait data from 37 healthy participants across 207 sequences.
• Employed artificial neural networks trained with data from a Vicon system for post-processing.
• Evaluated two training strategies: improving kinematic signals versus improving gait descriptors.

Main Results:

• Both supervised learning approaches demonstrated considerable accuracy improvements.
• Kinematic gait signals exhibited reduced errors and increased correlation with ground truth.
• Gait descriptor accuracy substantially improved, particularly for kinematic descriptors.
• No single training approach proved definitively superior.

Conclusions:

• Supervised learning effectively boosts 3D camera accuracy in gait analysis.
• The choice of training approach depends on specific study objectives.
• 3D cameras show significant potential for future gait analysis research.