An EMG-based Gesture Recognition for Active-assistive Rehabilitation

Physical disabilities affect many people worldwide, but therapies and rehabilitation can help one's recovery. Rehabilitation can be done manually. However, modern developments in technology enabled the use of robotics in rehabilitation, and has proved to be with comparable effectivity with manual rehabilitation. In line with this, a locally developed rehabilitation for upper limb extremities was developed. Regarding features, it can be at par with commercially available rehabilitation devices. To further improve the developed device, an algorithm to assist the user while rehabilitation is studied. Electromyography (EMG) based signals were used for input. Multiple models were compared, namely artificial neural networks (ANN), support vector machines (SVM), and recurrent neural networks (RNN). Six feature extraction techniques were used for training. These are integrated EMG (IEMG), root-mean-square (RMS), discrete wavelet transform (DWT), fast Fourier transform (FFT), waveform length (WL), and zero crossing (ZC). The best performing algorithm based on processing speed and training time was with IEMG as feature extraction technique, and ANN as training algorithm. Recorded training time was at 21 seconds, with the model's accuracy averaging at 96.3% and its processing time logged at 203 milliseconds. It was also interfaced with the existing rehabilitation device via simulation. Future directives include baseline compensation, correction techniques for muscle fatigue, and gathering data from healthy subjects and actual stroke patients.