System-Level Design of an End-to-End Neural Framework for Human Activity Recognition Using FMCW Radar

The growing demand for non-invasive sensing technologies has spurred increasing interest in radar-based gesture and activity recognition for applications in smart homes, healthcare, and human–computer interaction. Most existing approaches adopt a two-stage pipeline, combining traditional Fourier transform-based preprocessing with deep neural classifiers. In this work, we propose a complex-valued end-to-end neural architecture that directly processes raw frequency-modulated continuous wave (FMCW) radar signals, eliminating the need for handcrafted preprocessing. The proposed framework integrates a Fourier Multilayer Perceptron (Fourier MLP) with Fourier-based weight initialization to jointly learn multidimensional range, Doppler, and angle-of-arrival (AoA) representations while compressing redundant information through a dimensionality contraction layer. A Feature Enhancement Module based on complex convolutions further refines spectral–spatial representations by enhancing both high-and low-frequency components, and a Complex-Valued Self-Attention Encoder captures spatial–temporal dependencies across frames. Experimental results show that the proposed model achieves recognition accuracies of 98.54%, 99.58%, and 99.72% on hand gesture (HGR), arm gesture (AGR), and human activity (HAR) recognition tasks, respectively. Beyond its high accuracy and computational efficiency, the framework exhibits strong generalization to unseen users, thereby demonstrating superior robustness for real-world consumer applications.