Multiday Personal Identification and Authentication Using Electromyogram Signals and Bag-of-Words Classification Models

This article evaluates the performances of bag-of-words (BoWs) classification models on biometric applications using surface electromyographic (sEMG) signals generated by hand gestures. Extensive tests have been conducted on a publicly available multichannel multisession dataset collected from electrodes placed on the forearm and wrist of 43 persons while performing a set of 16 distinct gestures. FFT-based features extracted from six nonoverlapping frequency bands were combined with a BoW classifier and evaluated on authentication and identification tasks. A systematic ablation study considers the influence of the encoding strategy, the codebook dimension, and the length of the gesture-based password on the performances assessed in terms of the area under curve (AUC), equal error rate (EER), and cumulative match characteristics (CMCs). The definition of the training and test sets considered both within-day (WD) and cross-day scenarios. In the former case, average AUC and EER values indicate almost perfect operation for a password defined by three successive gestures, while CMC analysis showed Rank-5 performances above 99%. In the latter case, average AUC, EER, and Rank-5 CMC exhibited a small degradation of 1.1%, 3.1%, and 3.2%, respectively, showing significant robustness and improved performances against existing solutions.