A Real-time Hand Gesture Recognition System using 24 GHz Radar Array

This paper presents a description of a real-time hand gesture recognition system. This system consists of three commercial modules perpendicular mounted in an three-dimensional array to provide six-channel baseband I/Q signals. The I/Q signals are pre-processed by the doppler signal amplitude threshold detection and spectral analysis. A convolutional neural network consisting in two convolutional layers and two fully connected layers is constructed as the recognition classifier with less dependence of feature extraction. The network is trained with 1000 groups of datasets and verified by testing recognized results as the customized shortcut keys. Results show that this system could achieve a high recognition accuracy rate higher than 95% in the real-time test.

I. Introduction

Hand gesture recognition is an efficient way for Human Machine Interaction. The key problem in human gesture recognition is feature extraction and classification. Rather than cameral imaging and inertial sensors, radar has been suggested as an alternate method for recognizing hand gestures [1] . In this paper we use doppler radars to detect the velocity information of hand gesture and to reach a real-time control of a windows computer based on 24 GHz commercial doppler radars. The essential work of this project is to recognize the five hand gestures and later to implement the computer control as such as page up and down, window switch, and window close according to the previous recognition result. There are lots of developments about human gesture recognition using millimeter wave radars. Most impressive progress has been made, since 2016, by Google’s milestone work “Soli” using a 60 GHz Frequency Modulated Continuous Wave radar to fulfill hand recognition [2] . Our project aims to investigate the feasibility of recognizing different hand gesture based on commercial lost-cost radars with a three-dimensional array configuration. The classification is based on a convolutional neural network with the optimization in gradient descending and anti-over-fitting. The prototype has been developed and tested with higher accuracy than 95% in the real-time scenario. This result verifies the proposed method with a promising vision in the modern human-computer interactions.