Touchless hand gesture recognition is of great importance for human-computer interaction (HCI). In this paper, we present a hand gesture recognition approach based on range-Doppler-angle trajectory and the long short-term memory (LSTM) network with a 77GHz frequency modulated continuous wave (FMCW) multiple-input-multiple-output (MIMO) radar. Firstly, the hand gesture fast-time-slow-time-antenna 3 dimension (3D) data are collected by the FMCW MIMO radar. Additionally, by performing the discretize Fourier transform (DFT) to the fast-time and slow-time, respectively, we obtain the range-profile and Doppler-profile. Then, by using the multiple signal classification (MUSIC) approach, we estimate the angle-profile of the hand gestures. To smooth and eliminate the noise effects, we apply the Kalman filtering to the estimated range-profile, Doppler-profile and angle-profile, respectively, and obtain the range-Doppler-angle trajectory signature. After that, by exploiting the temporal and spatial correlations, we construct a LSTM network for the hand gesture recognition. Experiments with 6 hand gestures are conducted and show that the proposed approach can recognize 6 hand gestures with an average accuracy over 97%.
Recently, hand gesture recognition based-on radar has attracted many researchers in the field of human–computer interfaces. However, the number of kinds of hand gestures and recognition accuracy can be still increased. In this paper, we propose a hand gesture recognition approach based on convolutional neural network (CNN) using a bistatic radar system. Firstly, we build a bistatic radar system which consists of two pulse radars and define an active area of hand gesture called gesture desktop. Then, two time-distance maps are obtained by signal pre-processing, and we build a Bistatic-CNN with two branches of convolutional layers as a classifier to recognize 14 hand gestures. The bistatic radar system can offer us much more information of hand gesture from different perspectives and achieve much higher hand gesture recognition accuracy than single radar. The experimental results based on the measured data show that the proposed approach can recognize 14 hand gestures with average accuracy over 98%.