Improving Optical Myography via Convolutional Neural Networks

In order to improve the accuracy and reliability of myocontrol (control of prosthetic devices using signals gathered from the human body), novel kinds of sensors able to detect muscular activity are being explored. In particular, Optical Myography (OMG) consists of optically tracking and decoding the deformations happening at the surface of the body whenever muscles are activated. OMG potentially requires no devices to be worn, but since it is an advanced problem of computer vision, it incurs a number of other drawbacks, e.g., changing illumination, identification of markers, frame tear and drop. In this work we propose an improvement to OMG as it has been recently introduced, namely we relax the need of precise positioning and orientation of the markers on the body surface. The small size of the markers and their curvature while adhering to the surface of the forearm can lead to missed detections and misdetections in their orientation; here we rather detect the deformations by applying a Convolutional Neural Network to the region of interest around the feature source segmented, from the forearm. The classification-based approach yields results similar to those obtained by other classification based modalities, reaching accuracies in the range of 96.21% to 99.30% when performed on 10 intact subjects.