SRDST: Effective Dynamic Gesture Recognition With Sparse Representation and Dual-Stream Transformers in mmWave Radar

Abstract

Millimeter-wave radar holds significant potential for dynamic gesture recognition in contactless human-computer interaction, particularly in the Internet of Things and consumer electronics applications. However, a considerable challenge persists in filtering vast amounts of extraneous data from millimeter-wave radar echoes to isolate meaningful gesture features. We present a novel approach based on sparse representation principles to address this. We first generate a range-Doppler map of gestures using a two-dimensional (2-D) fast Fourier transform, then construct a Doppler-Time trajectory from aggregated data across multiple frames. Capitalizing on the intrinsic sparsity in the Doppler-time domain, we employ the orthogonal matching pursuit algorithm to refine a multidimensional feature sequence across time, Doppler, and range dimensions. Central to our approach is a dual-stream Transformer network that explores complex 2-D correlations in feature sequences via multihead self-attention mechanisms. This technique significantly improves gesture feature extraction efficiency and reduces data redundancy. The experimental results show that our model has an average recognition accuracy of 99.17% and a size of 0.17M, which is very suitable for application in embedded devices.