An efficient approach to recognize hand gestures using machine-learning algorithms

Electromyography (EMG) from a subject's upper limb can be used to train a machine-learning algorithm to classify different hand gestures. However, variability in the EMG signal due to between-subject differences can substantially degrade the machine-learning performance. This variation is usually due to the differences in both anatomical and physiological properties of the muscles, levels of muscle contraction, and inherent noises from the sensors. The aim of this study is to develop a subject-independent algorithm that can accurately classify different hand gestures. To minimize the between-subject differences, some selected time-domain EMG features are normalized to the area under the averaged root mean square curve (AUC-RMS). Five adult subjects with ages ranging 20–37 years performed three hand gestures including fist, wave-in, and wave-out for ten to twelve times each. Five machine-learning algorithms, including ¿-nearest neighbor (KNN), discriminant analysis (DA), Naïve Bayes (NB), Random Forest (RF), and Support Vector Machine (SVM) were used to classify the three different hand gestures. The results showed that the EMG features were moderately to strongly correlated with the AUC-RMS values. The SVM yielded maximum classification accuracy using the original EMG features (97.56%) which was significantly improved by using the normalized EMG features (98.73%) (p<0.05). The accuracy distribution of all classifiers were found to be closer to mean values when using the normalized EMG features compared to using the original EMG features. The developed approach of classifying different hand gestures will be useful in biomedical applications such as controlling exoskeletons and in certain human-computer interaction settings.