Automatic Grasp Selection using a Camera in a Hand Prosthesis.

Area of Science:

• Robotics
• Computer Vision
• Biomedical Engineering

Background:

• Current prosthetic hands often rely on electromyography (EMG) control, which can be limited in precision and object recognition.
• Integrating visual feedback into prosthetic control systems is an emerging area for enhancing functionality.

Purpose of the Study:

• To develop and evaluate an automatic grasp selection system for prosthetic hands using in-hand camera vision.
• To assess the accuracy and real-time applicability of a convolutional neural network (CNN) for classifying object grasps.

Main Methods:

• A camera was integrated into the palm of a prosthetic hand to capture object images.
• A labeled dataset was created using common graspable objects from ImageNet and custom-captured images.
• A convolutional neural network (CNN) was trained on this dataset for grasp classification.

Main Results:

• The CNN achieved a high grasp classification accuracy of 93.2%.
• Real-time demonstration showed successful grasp selection based on visual input.
• The system proved effective in augmenting existing EMG-controlled prosthetic hands.

Conclusions:

• Placing a camera in the palm of a prosthetic hand enables effective automatic grasp selection.
• CNN-based visual analysis significantly enhances prosthetic hand capabilities.
• This approach offers a promising avenue for improving the dexterity and usability of prosthetic devices.