Computer Vision and Deep Learning for Environment-Adaptive Control of Robotic Lower-Limb Exoskeletons.

Area of Science:

• Robotics and Artificial Intelligence
• Computer Vision and Deep Learning
• Human-Computer Interaction

Background:

• Robotic exoskeletons currently rely on manual control for locomotion mode switching, posing cognitive burdens on users.
• Automated locomotion mode recognition is crucial for enhancing the usability and intuitiveness of robotic exoskeletons.
• Existing datasets lack the diversity and scale required for robust environment recognition in real-world scenarios.

Purpose of the Study:

• To design and evaluate an environment recognition system for robotic exoskeleton control using computer vision and deep learning.
• To introduce and benchmark the ExoNet database, the largest open-source dataset for wearable camera images of diverse walking environments.
• To establish a baseline performance for automated environment recognition applicable to intelligent exoskeleton controllers.

Main Methods:

• Development of the ExoNet database, featuring diverse indoor and outdoor walking environments with hierarchical annotations.
• Training and testing of the EfficientNetB0 convolutional neural network, optimized via neural architecture search.
• Utilizing deep learning for image classification to predict walking environments from wearable camera data.

Main Results:

• The developed environment recognition system achieved an image classification accuracy of approximately 73% on the ExoNet database.
• This study provides the first benchmark performance metrics for the ExoNet dataset.
• The EfficientNetB0 model demonstrated effective environment prediction capabilities.

Conclusions:

• The proposed computer vision and deep learning system shows promise for enabling automated environment recognition in robotic exoskeletons.
• The ExoNet database serves as a valuable resource for advancing research in intelligent exoskeleton control.
• Further research is needed to explore various convolutional neural networks for real-time, adaptive locomotion mode recognition.