The recognition of multi-finger prehensile postures using LDA

The prehensile hand gestures play an important role in daily living for seizing or holding subjects stably. In order to realize the accurate recognition of eight prehensile hand gestures with a minimal number of electrodes, an off-line myoelectric control system with only two electrodes is developed. We choose Mean Absolute Value, Variance, the fourth-order Autoregressive Coefficient, Zero Crossings, Mean Frequency and Middle Frequency as original electromyography feature set and utilize the linear discriminant analysis to reduce the dimension and complete classification. The extent of dimension reduction is investigated and on the premise of it, the average accuracy can achieve 97.46% in the recognition of six hand gestures. The optimal feature set based on the original feature set is determined to be Mean Absolute Value, Variance, and the fourth-order Autoregressive Coefficient, which yields an average accuracy of 95.94% in the recognition of eight hand gestures. An average method is proposed to improve the accuracy further, resulting in the average accuracy in eight gestures being 98.12% and the best individual accuracy of some hand gestures being 100%.