On-Line Event-Driven Hand Gesture Recognition Based on Surface Electromyographic Signals

This paper presents a minimum complexity hand movement recognition algorithm based on Average Threshold Crossing (ATC) technique. It exploits the number of threshold-crossing events, generated by a full-custom acquisition board, from the surface ElectroMyoGraphic (sEMG) signals of three forearm muscles to detect four different movements of the wrist: flexion, extension, abduction and grasp. A Support Vector Machine (SVM) model has been trained with the signals acquired from ten subjects, who repeated ten times each gesture. To avoid correlation between training and testing dataset, the Leave One Subject Out (LOSO) cross-validation technique has been chosen. The average ATC classifier's accuracy is 92.87 %, only 5.34 % below the results obtained feeding the same model with the sEMG features extracted from the raw sampled signals. The total latency of the algorithm, from the acquisition to the prediction, is 160 ms. Power consumption was considered too: with less than the power budget for one sampled sEMG channel, it is possible to acquire and transmit (through a Bluetooth low energy module) the event-driven data of four sEMG channels, with an effective data rate of only 28 B/s. Obtained performance makes this technique suited for wearable systems or Internet-of-Things (IoT) applications.