Abstract
Gesture recognition is a topic in computer science with the goal of describing human gestures through mathematical algorithms in recent year. In the field of hand gesture recognition,it apply in many kinds of technologies such as mobile phone applications, wearable wireless devices, sports detection, video game or art combination. In this thesis, we will record signals of eight kinds of hand movements into computer using wearable wireless device with nine axis sensor (including accelerometer, gyroscope and magnetometer) worn on the wrist, then recognized gestures using the algorithms being described later. We built a system of recognition with machine learning classification process. Besides classification process, we also developed a thresholding method to easily detect movements. In the thresholding method, for each movement, we defined threshold value for each kind of data and filtered the movements data with threshold combined with detection windows. However, not all the movements can be detected by this easy and less calculation method so that we finally used a machine learning process to solve problems. The analyzing of the two method will be introduced later. In order to achieve higher recognition accuracy, we used machine learning process in the system and did feature extraction to get well distinguished features. We used principal component analysis (PCA) and linear discriminant analysis (LDA) to extract features. The advantages of PCA and LDA are reducing dimensions of data while preserving as much of the class discriminatory information as possible and reducing the training time of classification. Last, with support vector machine (SVM), we can recognize movement with higher accuracy with less computation time, and it also support data with high dimension. We can model even non-linear relations with more precise classification due to SVM kernels. In our experiment, we can get the accuracy of recognition at 99.63% for 8 classes with 20 subjects data for 5 times each in user-dependent case, and 12 subjects testing data for user-independent case with recognition rate at 88.43%.