A lightweight network with multi-feature fusion for mmWave radar based hand gesture recognition

In this article, we consider the problem of hand gesture recognition (HGR) using millimeter-wave (mmWave) radar. With the aim of improving the HGR performance while using low computational complexity and storage resource, a lightweight network with multi-feature fusion is proposed, which extracts and fuses the features from the range-time maps (RTMs) and angle-time maps (ATMs). Specifically, the input layer is applied to input and fuse the RTMs and ATMs. Then, the lightweight units are designed to extract features with little computational complexity. After that, the channel attention module is used to learn the important parts of the features. Finally, the classification layer is employed to output the predicted hand gesture results. Experimental results on real data show that the accuracy of the proposed method reaches 97.53% in the conditions of 0.92M Params and 0.100G FLOPs, which verifies the effectiveness and reasonableness of the proposed method.