Development and comparison of dataglove and sEMG signal-based algorithms for the improvement of a hand gestures recognition system.

Hand gesture recognition is a topic widely discussed in literature, where several techniques are analyzed both in terms of input signal types and algorithms. The main bottleneck of the field is the generalization ability of the classifier, which becomes harder as the number of gestures to classify increases. This project has two purposes: first, it aims to develop a reliable and high-generalizability classifier, evaluating the difference in performances when using Dataglove and sEMG signals; finally, it makes considerations regarding the difficulties and advantages of developing a sEMG signal-based hand gesture recognition system, with the objective of providing indications for its improvement. To design the algorithms, data coming from a public available dataset were considered; the information were referred to 40 healthy subjects (not amputees), and for each of the 17 gestures considered, 6 repetitions were done. Finally, both conventional machine learning and deep learning approaches were used, comparing their efficiency. The results showed better performances for dataglove-based classifier, highlighting the signal informative power, while the sEMG could not provide high generalization. Interestingly, the latter signal gives better performances if it’s analyzed with classical machine learning approaches which allowed, performing feature selection, to underline both the most significative features and the most informative channels. This study confirmed the intrinsic difficulties in using the sEMG signal, but it could provide hints for the improvement of sEMG signal-based hand gesture recognition systems, by reduction of computational cost and electrodes position optimization.