Radar-Based Hand Gesture Recognition With Feature Fusion Using Robust CNN-LSTM and Attention Architecture

In Human-Computer Interaction (HCI), seamless hand gesture recognition is essential for intuitive and natural interactions. Gestures act as a universal language, bridging the gap between humans and machines. Radar-based recognition surpasses traditional optical methods, offering robust interaction capabilities in diverse environments. This article introduces a novel deep learning approach for hand gesture recognition, leveraging convolutional neural networks (CNNs), long short-term memory networks (LSTMs), and attention mechanisms. CNNs extract spatial features from radar signals, while LSTMs model the temporal dependencies crucial for dynamic gestures. Additionally, attention mechanisms enhance feature selection, ultimately improving recognition performance. We evaluate our method on the UWB-Gestures dataset, which has 12 gestures from eight people that were recorded using three X-Thru X4 UWB impulse radar sensors. Our processing pipeline integrates feature extraction, LSTM-attention blocks, and dense layers for final classification. Early fusion techniques, which combine spatial and temporal features in the initial stages, yield superior results, achieving an overall accuracy of 98.33% and outperforming intermediate fusion methods across gesture classes. To enhance the model’s robustness, we evaluated its performance under common contributors of radar-specific noise scenarios in practical applications, including Gaussian noise, signal inversion, and multipath interference. Our model demonstrates high resilience, maintaining performance despite adverse conditions. As compared to state-of-the-art approaches, our approach delivers competitive accuracy and enhanced robustness, offering a reliable solution for noise-resilient radar-based hand gesture recognition in real-world applications.