Using Sparse Convolutional Networks for Micro-Doppler Signatures

Micro-Doppler signatures, which encode subtle motions such as human limb movements or drone rotor rotations, are pivotal for radar-based target recognition. However, traditional convolutional neural networks (CNNs) face challenges in handling the inherent sparsity and long-term temporal dependencies of such data. This paper proposes SC-Transformer, a hybrid framework integrating Sparse Convolutional Networks (SCN) and Transformer architectures. SCN reduces computational redundancy by focusing on non-zero regions in sparse velocity-time maps, while the Transformer captures global temporal dependencies via multi-head self-attention. Experimental results on a six-class human activity dataset demonstrate that SC-Transformer achieves 88.1% accuracy, outperforming ResNet-50 and CNN-LSTM with 60% fewer parameters. Time-series data augmentation strategies (e.g., temporal flipping, amplitude inversion) further enhance generalization under limited data conditions. This work provides a lightweight, efficient solution for radar-based target classification and advances sparse time-series modeling.