LOW COST MULTI-SENSOR GESTURE RECOGNITION GLOVE WITH DEEP LEARNING ON DEVICE PROCESSING

Abstract

This Master’s thesis puts together core aspects of gesture recognition, discusses its applications, reviews the performance of previously derived solutions, and then provides a new hardware/software framework for accurate recognition. Gesture recognition is a process of interpreting hand gestures. It can take advantage of the natural flexibility of human hands to provide a feasible alternative for human-computer interaction (HCI). Such gesture-based devices form only a minor part of input devices in general use, although they can offer improved performance, comfort, and creativity in particular tasks. The applications of gesture recognition vary from narrow medical rehabilitation and military planning to business, engineering, and personal use cases. State-of-the-art solutions lack necessary performance even for the general public with leisure activities, they also tend to concentrate on covering a limiting case of a single type of sensor. This research proposes the design of a low-cost glove in terms of hardware and a software method to tackle this subset of problems in gesture recognition. It opens further possibilities for new research in the area of gesture recognition. The resulting glove provides an improvement over previously defined research in terms of gesture recognition accuracy, as well as potentially cross-user applicability at low cost. The accuracy level on 31 different commonly utilized gestures is reaching 95-96%. It is suggested that a mathematical model is suitable for better since some of the gestures are vaguely defined and cause indiscernibility. Future work includes larger sets of test subjects, correlational analysis of sensory data and the use of other, perhaps, cheaper sensors with custom hardware.