Machine learning techniques for surface electromyography based hand gesture recognition

People use their hand unconsciously to perform many task in daily life. The goal of these tasks can be a physical interaction such as holding, grasping and pushing as well as communication. Hand gestures constitute one of the most popular types of non-verbal communication. People use their hands to give basic messages such as 'stop', 'come', 'like', 'dislike' and 'okay'. Even when texting, we now use hand gesture emojis. Moreover, hearing impaired people use hand gestures as sign language, which means hand gestures can be considered a language on their own. On the other hand, the communication does not take place only between humans, but also between humans and technological devices. Input method of these devices is generally limited with mouse, keyboard, joyistick or touchpad. The human machine interaction studies focus on developing new interfaces to overcome this restriction by simulating a natural ways of communication for humans. Here, the hand gestures arise as an interface for devices. There are various methods for hand gesture detection such as cameras, gloves and electrodes. Visual based solutions provide high recognition rate; however, they face with difficulties like light conditions, skin and background color, spatial position and orientation of the hand and fingers. Also, this kind of systems suffer from range limitations of sight view. Next, glove based systems hold people back from using their hands comfortably and may also cause loss of sensation. Hence, an alternative solution has been by generated electrode based systems. When contracted to perform an action, muscle cells generate an electrical signal. The combination of all innervated muscle cells produces superposed electrical signals called Motor Unit Action Potentials. A technique of recording and studying the myoelectric signal is defined as electromyography (EMG). This thesis mainly proposes a hand gesture recognition system using the electromyography signals of the forearm. The muscle signals can be recorded via needle electrodes or surface electrodes. While needle electrodes are providing signals from a specific muscle group, surface electrodes sense signals that reach the surface of skin. The use of non-invasive electrodes is painless and easy; therefore, they are known as convenient for daily usages and applications. In this work, a commercial product, the Myo Armband has been used as a signal source. The Myo Armband is a wearable and wireless device, which has eight electrodes located circularly around the arm. When a user wears the Myo Armband, the device transmits the measured EMG signal via Bluetooth and the signals can be easily obtained with help of software development kit of device. Hence, it is widely used in researches and applications like prosthetic control, mobile robot navigation, virtual reality projects, game control, sign to speech translation and art works. In this study, the hand gestures of numbers between 0 and 9 in Turkish Sign Language was recorded as a dataset using the Myo Armband. To make a better gesture recognition, the proposed system includes the comparison of four different machine learning algorithms which consist of k-Nearest Neighbor, Decision Tree, Support Vector Machines and Artificial Neural Network. In machine learning, working with directly raw data increases the complexity of implementation and processing cost. Thus, first, seven time domain features were extracted from each channel of the recorded data. To make the feature matrix simple, Principal Component Analysis was employed, and then the algorithms was fed by resultant feature vector. Before the comparison of algorithms, each algorithm was investigated to find the optimal parameters by changing the values. Firstly, k-Nearest Neighbor algorithm was examined with various number of neighbors and distance metrics. The Decision Tree algorithm was tested for two splitting criteria which are gini diversity index and entropy. The Support Vector Machine algorithm was analyzed with different kernels and binary discrimination methods. Lastly, Artificial Neural Network was investigated by changing the number of neurons in hidden layer, training function and error functions. Then the outcome of this study has obtained with setting the parameters of each classifiers to optimal. According that, Decision Tree algorithim has lowest fscore ratio which was about 0.76. k-Nearest Neighbor and Artificial Neural Network algorithms increased the result to 0.82 and 0.85, respectively. The highest fscore ratio was achieved as 0.87 using the Support Vector Machine algorithm. In addition, the performance of each class has also evaluated, and it was found that the Support Vector Machine algorithm gives the highest ratio for all gestures. Also, gestures of '6', '7', '8' and '9' have higher recognition ratio; whereas, lowest ratio was obtained for class '2'. The achieved results are quite promising that, the usage of muscle signals could employ as hand gesture recognition and the proposed system can be used to interact with machines.