A Gaussian Mixture Classification Model with Unlearned-class Detection for FPGA Implementation and Application for Classification of Combined Motions

Abstract

The electromyogram (EMG) signal generated by muscle contraction has been widely utilized for motion estimation of arms and fingers. To develop a myoelectric prosthetic hand that has high general versatility and safeness, a classifier that can consider complex forearm motions and motions that are not assumed during training, is required. However, hardware implementation of complex classifiers that has high classification performance is difficult. To tackle these problems, this paper proposed a novel probabilistic neural network that can be implemented in FPGA (Field Programmable Gate Array), and it was applied to an EMG-based human-machine interface system. The proposed neural network includes two types of probability density functions optimized for hardware implementation and enabled the execution of multi-class discrimination considering the unlearned class on the FPGA. Furthermore, by combining a forearm motion classifier and a hand motion classifier, the consideration of compound motions consisting of multiple hand gestures can be achieved. In experiments, the results showed that the proposed method can be implemented on FPGAs, and demonstrated that it can achieve highly accurate motion classification for compound motions and unlearned motions.