Vision for Prosthesis Control Using Unsupervised Labeling of Training Data

Abstract

Transitioning from one activity to another is one of the key challenges of prosthetic control. Vision sensors provide a glance into the environment’s desired and future movements, unlike body sensors (EMG, mechanical). This could be employed to anticipate and trigger transitions in prosthesis to provide a smooth user experience. A significant bottleneck in using vision sensors has been the acquisition of large labeled training data. Labeling the terrain in thousands of images is labor-intensive; it would be ideal to simply collect visual data for long periods without needing to label each frame. Toward that goal, we apply an unsupervised learning method to generate mode labels for kinematic gait cycles in training data. We use these labels with images from the same training data to train a vision classifier. The classifier predicts the target mode an average of 2.2 seconds before the kinematic changes. We report 96.6% overall and 99.5% steady-state mode classification accuracy. These results are comparable to studies using manually labeled data. This method, however, has the potential to dramatically scale without requiring additional labeling.