Abstract

This thesis presents a wearable, smart ring with an integrated Bluetooth low-energy (BLE) module. The system uses an accelerometer and a gyroscope to collect fingers motion data. A prototype was manufactured, and its performance was tested. To detect complex finger movements, two rings are worn on the point and thumb fingers while performing the gestures. Nine pre-defined finger movements were introduced to verify the feasibility of the proposed method. Data pre-processing techniques, including normalization, statistical feature extraction, random forest recursive feature elimination (RF-RFE), and k-nearest neighbors sequential forward floating selection (KNN-SFFS), were applied to select well-distinguished feature vectors to enhance gesture recognition accuracy. Three supervised machine learning algorithms were used for gesture classification purposes, namely Support Vector Machine (SVM), K-Nearest Neighbors (KNN), and Naive Bayes (NB). We demonstrated that when utilizing the KNN-SFFS recommended features as the machine learning input, our proposed finger gesture recognition approach not only significantly decreases the dimension of the feature vector, results in faster response time and prevents overfitted model, but also provides approximately similar machine learning prediction accuracy compared to when all elements of feature vectors were used. By using the KNN as the primary classifier, the system can accurately recognize six one-finger and three two-finger gestures with 97.1% and 97.0% accuracy, respectively.