Dynamic gesture recognition model based on FMCW radar and multidimensional feature fusion with dual-channel CNN and GRU

Abstract

With the development of computers, gestures have also become a commonly used human-computer interaction, so it becomes important to recognize gestures correctly. The objective of this paper is to propose a deep learning model to improve the accuracy of low performance radar in dynamic gesture recognition.We use a dataset acquired by CAL60S244-AB radar to classify ten gestures. The model consists of a dual-channel convolutional neural network(CNN), Gate Recurrent Unit(GRU) and a classification module. The inputs to the model are feature vectors fused in four dimensions of distance, speed, receiving antenna channel and time, from which the dual-channel CNN can extract shallow and deep spatial information, and GRU can extract temporal information, and the combination of these two parts makes the model's classification accuracy improve. Through experiments, the accuracy of the model on the dataset reaches 95.1%, and the results show that the model proposed in this paper can improve the accuracy of dynamic gesture classification for low-precision radar, which verifies the effectiveness of the model.