Spectral Domain Neural Reconstruction for Passband FMCW Radars

Harshvardhan C. Takawale, Nirupam Roy

Keywords: Implicit Neural Representation, Neural Volumetric Reconstruction, Frequency Modulated Continuous Wave, mmWave radar

Abstract: We present SpINR, a novel neural framework for high-fidelity volumetric reconstruction of small, near-field tabletop objects using Frequency-Modulated Continuous-Wave (FMCW) radar. Traditional radar imaging techniques often apply FFT as a black-box post-processing step, discard phase information, and require dense aperture sampling, leading to limitations in sub-centimeter reconstruction and generalization. Our core contribution is a fully differentiable frequency-domain forward model that captures the complex radar response using closed-form synthesis, paired with an implicit neural representation (INR) for continuous volumetric scene modeling. Unlike time-domain baselines and magnitude-only spectral pipelines, SpINR directly supervises the complex frequency spectrum from raw 1D chirps acquired along an arbitrary cylindrical aperture, preserving phase-sensitive spectral spectral fidelity while drastically reducing computational overhead. Additionally, we introduce sparsity and smoothness regularization to disambiguate sub-bin ambiguities that arise at high carrier frequencies and fine range resolutions. Experimental results show that SpINR significantly outperforms both classical and learning-based baselines, with a 52.6\% improvement in reconstruction quality and 32\% improvement in latency.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 14183