Robust UWB Radar Gesture Recognition Addressing Speed-Induced Scale Variations via Multi-Scale Feature Extraction

Abstract

In this paper, we propose a speed-robust hand gesture recognition system incorporating a multi-scale feature extraction (MSFE) module to address the critical engineering challenge of time-frequency scale mismatch caused by varying gesture speeds in ultra-wideband (UWB) radar. Conventional convolutional neural networks (CNNs) utilizing fixed kernels fail to effectively capture features when the same gesture is performed at different speeds, leading to performance degradation. To overcome this, our MSFE is specifically designed for the range-Doppler domain and is applied to the first layer to normalize physical motion-induced scale variations early in the pipeline. Experimental results on the UWB-gestures dataset demonstrate a 98.12% accuracy, outperforming conventional CNN and LSTM-based models. Furthermore, we provide a comprehensive runtime analysis, confirming that the proposed system is computationally efficient with only a marginal 1.3% increase in parameters, making it feasible for real-time deployment on edge devices.