Vision-aided grasp classification: design and evaluation of compact CNN for prosthetic hands

Area of Science:

• Robotics and Artificial Intelligence
• Biomedical Engineering
• Computer Vision

Background:

• Upper limb amputees require advanced prosthetic hands with diverse grasp capabilities.
• Accurate grasp pattern identification is essential for intuitive prosthetic control.
• Vision-based techniques using deep learning, like Convolutional Neural Networks (CNNs), offer non-contact control but face generalization challenges.

Purpose of the Study:

• To develop a compact CNN model (GraspCNet) for effective grasp classification in prosthetic hands.
• To enhance the model's ability to generalize across various object shapes, including unseen objects.
• To enable real-time grasp classification suitable for embedded prosthetic systems.

Main Methods:

• Proposed GraspCNet, a compact CNN utilizing separable convolutions to reduce computational load.
• Trained and tested the model on standard object datasets employing a cross-validation strategy.
• Evaluated performance on both seen and unseen object classes, and in computer-based real-time experiments.

Main Results:

• Achieved average accuracies of 82.22% for seen object classes and 75.48% for unseen object classes.
• Demonstrated 69% accuracy in computer-based real-time experiments.
• Outperformed most benchmark techniques and showed comparable results to DcnnGrasp.

Conclusions:

• GraspCNet effectively classifies grasp patterns, demonstrating strong generalization capabilities for prosthetic hand applications.
• The model's compact design is suitable for real-time embedded systems.
• GraspCNet shows potential for integration with other sensing modalities in advanced prosthetics.