Mechanomyogram Based Individual Finger Gesture Recognition System

Mechanomyograph(MMG) based hand gesture recognition technology is a kind of new-style human-machine interface technology which can be applied in the rehabilitation area,such as prosthetic control and rehabilitation training based on virtual reality. Thus, it is widely investigated in recent decades. However, among all the existing studies, none of them can recognize the individual movements of five fingers. In this thesis, a MMG based finger gesture recognition(MFGR) system is designed to identify the motions of individual finger.The MFGR system is explained in two parts in this thesis: one is the MFGR hardware design,and the other is the MFGR signal processing algorithms.For the design of the hardware, the MFGR system consists of three sub-systems: signal acquisition system, wireless communication system and signal processing system. In order to improve the wearability of the signal acquisition system, the MEMS based inertial sensors are used for the MMG signal detection. For the wireless communication system, the bluetooth low energy(BLE) technology is used. The signal processing system is designed and implemented in the Android platform as the Android system has strong computational capability and is widely used in daily life.For the MFGR signal processing part, to extract the finger movement events from the raw MMG signal automatically, a tapping event detection(TED) algorithm is designed. This algorithm is beneficial to the real-time data processing since it detects the finger movement faster and easier than the manual method which is normally adopted in the existing such kind of applications. In the field of pattern recognition, although a larger number of feature extraction and recognition algorithms have been proposed, few are used for the finger gesture recognition. Therefore, many kinds of pattern recognition algorithm, i.e. the wavelet packet transform(WPT), single value decomposition(SVD), support vector machine(SVM)classifier, Na?ve Bayes Classifier(NBC), etc., are tried for recognition in this thesis.The experimental results show that the MFGR algorithm is able to achieve an averageaccuracy rate of 93.1%. And for the MFGR system, the average recognition accuracy rate of81.3% is obtained in this thesis. Possible factors affecting the recognition performance are discussed in the end.