Sparsity-Based Range-Velocity-Time PointMLP for FMCW Radar Human Activity Classification with Efficient Computation

Abstract — Frequency Modulated Continuous Wave (FMCW) radar enables privacy-preserving human activity recognition (HAR), and most existing approaches rely on time-Doppler (TD) and range-Doppler (RD) spectrograms for classification. This work proposes a sparsity-driven formulation in which TD/RD sequences are converted into compact TDSP/DRDSP point sets using Fast Iterative Shrinkage-Thresholding Algorithm (FISTA) ℓ1 recovery. Eliminating the need for dense image formation, these sparse features are processed using an RVT-PointMLP architecture, which incorporates radar-aligned range-velocity-time metrics and integrates geometric affine normalization. Experiments on independent 5.8 GHz and 24 GHz datasets show that lightweight RVT-PointMLP models achieve 97.89% and 89.11% accuracy, respectively. These results indicate that sparsity-based point-cloud learning provides an efficient and scalable representation for radar HAR, supporting deployment in lightweight or real-time systems.