Surface Electromyography Signal Recognition Based on Deep Learning for Human-Robot Interaction and Collaboration

The interaction between humans and collaborative robots in performing given tasks has aroused the interest of researchers and industry for the development of gesture recognition systems. Surface electromyography (sEMG) devices are recommended to capture human hand gestures. However, this kind of technology raises significant challenges. sEMG signals are difficult to acquire and isolate reliably. The creation of a gesture representative model is hard due to the non-explicit nature of sEMG signals. Several solutions have been proposed for the recognition of sEMG-based hand gestures, but none of them are entirely satisfactory. This study contributes to take a step forward in finding the solution to this problem. A sEMG capturing prototype device was used to collect human hand gestures and a two-step algorithm is proposed to recognize five valid gestures, invalid gestures and non-gestures. The former algorithm step (segmentation) is used for sEMG signal isolation to separate signals containing gestures from signals containing non-gestures. The latter step of the algorithm (recognition) is based on a deep learning method, a convolutional neural network (CNN) that identifies which gesture is in the sEMG signals. The performances of the prototype device and recognition architecture were compared successfully with the off-the-shelf sEMG device Myo. Results indicated that the segmentation process played an important role in the success of the gesture recognition system, excluding sEMG signals containing non-gestures. The proposed system was applied successfully in the control loop of a collaborative robotic application, in which the gesture recognition system achieved an online class recognition rate (CR) of 98%, outperforming similar studies in the literature.