Gesture Recognition Using LFMCW Radar and Convolutional Neural Network

Abstract

With the continuous enrichment of human-computer interaction, gesture recognition technology plays an increasingly important role in various fields, especially in the non-contact and low-light conditions. In order to overcome the limitations of traditional visible optical sensors, such as low-light, occlusion and privacy, a gesture recognition system based on linear frequency modulation continuous wave (LFMCW) radar is proposed. By preprocessing the LFMCW radar intermediate frequency (IF) signals, the time series of the two-dimensional range-doppler(R-D) maps of the gesture is obtained. The R-D map series is taken as the three dimensional convolutional neural network (3D-CNN) input samples, and the gesture features are extracted via convolution layer and pooling layer, and the gesture features are classified and identified via the fully connected soft-max classifier. To explore the application of the extracted gesture features, a 77GHz radar system is applied, and six kinds of gestures are designed to verify the performance of classification and recognition. The experimental results show that the gestures can be classified and recognized effectively and the real-time recognition accuracy of the whole system can reach about 94%.