Effect of Number of Features on sEMG Multi-sensor Combinations for Hand Gesture Recognition

Abstract

In literature, it is common to use multiple wearable sensors to capture the characteristics of a hand gesture in applications requiring hand activity classification. Using machine learning, the design of hand gesture recognition systems can be optimized in terms of the selection of relevant number of sensors and features. Hence, the computation complexity, power consumption and hardware requirement of the system may be reduced while the classification accuracies are enhanced. In this paper, five surface electromyogram (sEMG) sensors are placed on the forearm muscles. A set of 20 activities is considered, which constitute five different hand postures each recorded for two distinct orientations and two distinct hand motions. The classification accuracies are determined when number of sensors and number of relevant features for each multi-sensor combination are varied. Results indicate that, as the number of relevant features is increased, the classification accuracies improve although the improvement becomes less significant after a certain number of features is available for classification. Also, the best classification accuracies obtained using 4 or 5 sEMG sensors are comparable having a value of 96.2% and 95%, respectively. There is a slight reduction in the average classification to 92.7% when 3 sEMG sensors are used. However, using 2 or just one sEMG sensor the degradation in the classification accuracies is significant, with a value of 89.3% and 65.6%, respectively, even when all the relevant features are considered for classification. Hence, the minimum number of sensors required for carrying out classification of the considered hand gestures is determined to be three, since they provide the best tradeoff in terms of quality and cost. The statistical significance of the change in the classification accuracies with the change in number of sensors and the number of features is checked using analysis of variance (ANOVA).