Spiking Convolutional Neural Network for Hand Movement Classification Using SEMG and Kinematic Data

Abstract

Gesture recognition using surface Electromyography (sEMG) data is an emerging field of study that can help enable reliable human-machine interaction, with applications in the development of advanced prostheses. Despite Machine Learning and Deep Learning approaches show promising outcomes in classification of sEMG signals, most efforts still focus on feature extraction and data preprocessing; furthermore, the models involved bear complex architectures that require advanced preprocessing techniques, while their implementation on portable devices can be challenging. Spiking Neural Networks (SNNs) offer a biologically inspired computational paradigm. Its event-driven nature makes it suitable for processing temporal data; moreover, its implementation at the edge using certain hardware platforms, such as Loihi requires a low energy budget. This paper proposes a Spiking Convolutional Neural Network (SCNN) for the classification of 12 gestures using the NinaPro DB5 Dataset, where the sEMG signals were upsampled by linear interpolation and encoded using delta algorithm. The performance achieved yields 96% and 80.55% for training and test accuracy, respectively, and the inference time accomplished is 5.51s.