Development of A Learning-Based Terrain Classification Framework for Pushrim-Activated Power-Assisted Wheelchairs^.

Area of Science:

• Assistive Technology
• Robotics
• Machine Learning

Background:

• Pushrim-activated power-assisted wheels (PAPAWs) offer torque assistance but can face control issues due to environmental changes.
• Current PAPAW controllers lack adaptability, leading to inefficient or unsafe wheelchair use.
• Adaptive control strategies are needed to enhance PAPAW safety and stability by recognizing terrain.

Purpose of the Study:

• To develop and evaluate a learning-based terrain classification framework for PAPAWs.
• To create a context-aware sensory system for recognizing diverse indoor and outdoor terrain conditions.
• To enable adaptive velocity/torque control for improved wheelchair performance.

Main Methods:

• Collected sensor data from wheelchair-mounted gyroscopes and accelerometers during various propulsion routines.
• Extracted relevant features from sensor measurements.
• Trained and tested machine learning classifiers for terrain identification, comparing one-stage and two-stage approaches.

Main Results:

• A one-stage multi-label classification framework demonstrated superior accuracy over a two-stage pipeline.
• Outdoor terrains were classified with higher average accuracy (90%) compared to indoor terrains (65%).
• The proposed framework effectively distinguishes between different environmental conditions for PAPAWs.

Conclusions:

• The developed learning-based framework enables real-time terrain classification for PAPAWs.
• This classification provides crucial information for designing adaptive velocity/torque controllers.
• The system has the potential to significantly improve wheelchair maneuverability and user safety.