An Efficient Deep Learning Approach for sEMG Hand Gesture Classification

Amputees face psychological and sensory issues in their daily lives due to missing essential human activities and shape. Fortunately, prosthetic solutions can reduce the gap between amputees and normal people's life quality by providing human-like limbs that are able to perform important movements. This paper proposes a highly accurate sEMG hand and wrist gesture classifier that can be used for real-time applications with prosthetic arms, focusing on the preprocessing of raw recorded sEMG signals and the development of a Convolutional Neural Network (CNN) classifier. sEMG is a very popular way to record electrical muscle activity due to its simplicity and safety. CNN has been proven to be an efficient Machine Learning method in classification tasks and we further improve its ability by proposing a simple but effective preprocessing technique using Fast Fourier Transform and overlapping Windows input to a simple CNN model. The presented model is able to successfully classify 17 different hand and wrist gestures with an accuracy of up to 99.64% from sEMG signals collected from 10 subjects used from the NinaPro database5 using CNN.