Machine learning techniques for hand gesture recognition using short range contactless microwave sensor
Thomas, Ashita Priya
Date of Issue2018
School of Electrical and Electronic Engineering
Hand gesture recognition is a topic of interest for many researchers and business sectors and various methodologies to perform gesture recognition are still being explored. Hand gestures play a major role in the interaction and control of devices/machines in applications such as home appliances, automated systems, gaming, virtual reality, etc. The recognition of fine movements of fingers in the hand with high precision is a challenging task. The existing technologies can sense the motion of human body or hand in a large scale using several input devices such as depth-cameras, IR sensors, etc. The latest advancement is the development of hand gesture recognition techniques through RF sensing, but this requires a complex feeding mechanism to drive the beam scanning of antenna and the field of view for scanning is not wide enough. A method which incorporates a simple feeding mechanism, wide field of view and small number of circuitry to recognize hand gestures can be very useful for interacting with devices, especially used in consumer electronics applications. In this work, a study on hand gesture recognition techniques using microwave radar system has been done. Also, a novel method of using a continuous wave radar system employing a 10 GHz leaky wave antenna for gesture recognition is proposed in this work. Two approaches have been proposed in this work. The first approach makes use of a single leaky wave antenna which acts as a transmitter as well as a receiver for recognizing three different hand gestures. The drawback of this approach is that the gesture data obtained from the sensor is found to be ambiguous due to the superimposition of the transmitted and received signal. Hence a second approach is used, where two leaky wave antennas (one as transmitter and the other as receiver) are used to sense four different hand gestures. Useful features such as median, rms, etc. are extracted from the data. Machine learning algorithms such as Random Forest, Support Vector Machine and Linear Support Vector Machine have been used for classifying these gestures and an average accuracy of 96.43 %, 94.05 % and 96.43 % are achieved respectively. The classification has been done in both Weka (a data mining tool) as well as Python and a comparison of different machine learning algorithms with the chosen three algorithms has been done.
DRNTU::Engineering::Electrical and electronic engineering