Realtime recognition of multi-finger prehensile gestures

Abstract

This paper presents an on-line myoelectric control system which can classify eight prehensile hand gestures with only two electrodes. The overlapping windowing scheme is adopted in the system leading a continuous decisions flow. We choose mean absolute value (MAV), variance (VAR), the fourth-order autoregressive (AR) coefficient and Sample entropy (SampEn) as the feature set and utilize the linear discriminant analysis (LDA) to reduce the dimension and obtain the projected feature sets. The current projected feature set and the previous one are “pre-smoothed” before the classification, and then a decision is generated by LDA classifier. To get the final decision from the decisions flow, the current decision and m previous decisions are “post-smoothed”. The method mentioned above can obtain a 99.04% off-line accuracy rate and a 97.35% on-line accuracy rate for individual gesture. By choosing a proper value of m, this method can also get a 99.79% accuracy rate for on-line recognition of complex sequences of hand gestures without interruption. In addition, a virtual hand has been developed to display the on-line recognition result visually, and a proper control strategy is proposed to realize the continuous switch of hand gestures.