ML-HGR-Net: A Meta-Learning Network for FMCW Radar Based Hand Gesture Recognition

In recent years, radar-based hand gesture recognition (HGR) has attracted much attention in the field of human-computer interaction (HCI) due to its benefits of high recognition accuracy and independence from lighting conditions. Conventional deep learning (DL) based models for HGR rely on a large amount of labelled data for training to achieve a satisfactory recognition accuracy. However, it is usually time-consuming and labor-intensive for collecting large amounts of radar data and also difficult to cover all possible hand gesture classes. In this paper, we introduce meta-learning to address the few-shot learning problem for frequency modulated continuous wave (FMCW) radar based HGR. We propose a meta-learning network to learn a model that can quickly adapt to unseen hand gesture tasks with few training observations. In particular, we propose a 3D convolutional neural network (CNN) based dual-channel fusion network for feature extraction by exploiting the correlations of multiple features in radar echo signals to improve the recognition accuracy. In addition, we also develop a learnable relation module with neural networks as a non-linear classifier to measure the similarity between the samples of different hand gestures, which can be more applicable for the HGR task. Finally, we evaluate the performance of the proposed model by conducting experiments on an FMCW radar hardware system. Experimental results show that the proposed meta-learning model substantially enhances the recognition accuracy compared with the state-of-the-art methods including DL and meta-learning based models for FMCW radar-based HGR.