Hand Gesture Signal Classification using Machine Learning

This research work focuses on identifying a specific hand gesture from the given EMG signal, acquired by sensor-based band. Surface EMG and machine learning techniques are used for the identification and classification purpose. The raw EMG signal captured using the sensor is initially passed through suitable preprocessing steps to avoid the noise artifacts. Followed by this, 8 different time-domain features are collected from these raw EMG signals, using which a feature matrix is created. SVM and KNN are the machine learning classifiers used here. The entire system is implemented in MATLAB 2019a. Using these methods, a promising accuracy of 93% is obtained through KNN and an accuracy of 83% using SVM.

I. Introduction

Hand gesture recognition is an area of active research in the computerized world as well as in language technology which is performed using mathematical algorithms. Simple gesture movement recognition plays a significant role in human-computer interaction which enables one to efficiently control the devices. Hand gesture recognition methods are more accurate, highly stable and take little time to control devices. Hand gestures can be collected in various forms such as images, graphs or numerical values. When the human hand poses for a specific hand gesture the muscle fibers will get a contract and thereby produces a small electrical signal. These signals are collected by electrodes. There are different modes of hand gesture collection, by capturing images, by using gloves, and by using electrodes. On using the camera the input is in the form of image captured with a high recognition rate but still, it faces some problems such as the background light, skin, orientation of the fingers, etc. The glove system may not be comfortable to use and hence we could not obtain the input data as we need. Hence the only alternative solution is the electrode system. Here we collect data using electrodes. For that, we are using a Mio armband which consists of 8 electrodes. The EMG signals are collected using these electrodes and data values are collected from each channel corresponding to the time in ms. These data are taken from a publically accessed dataset called Mio dataset where seven hand gestures from 36 able bodies subjects are collected. These are available in the form of numerical signal values and with this, the whole process is done. Here in our project, we aim to recognize the hand gesture corresponds to the given EMG signal. The dataset for this is taken from the Mio dataset and the Ninapro dataset. From this, the training and testing datasets are then divided. The data from the Mio dataset is used for training purposes and for testing we took data from the Ninapro dataset. The data in these datasets are in the form of signals which are collected using hardware called Myo armband, which can be easily worn on hands while recording the gesture. Here we explained the different preprocessing steps, feature extraction methods, and classification techniques.