A Physics-Informed Digital Twin Framework for Calibrated Sim-to-Real FMCW Radar Occupancy Estimation

TL;DR

This work tackles the challenge of obtaining robust radar perception models without extensive real-world labeling by introducing a physics-informed digital twin and calibrated domain randomization (CDR) for sim-to-real transfer in FMCW radar occupancy detection and people counting. The method builds a digital twin using SBR and Physical Optics to generate RD maps from a 60 GHz FMCW radar, paired with real unlabeled calibration data to align noise-floor statistics while preserving micro-Doppler cues. Empirical results show that a ResNet18 trained with CDR-adjusted simulated data achieves 97% occupancy accuracy and 72% counting accuracy on real data, outperforming ray-tracing baselines and uncalibrated domain randomization. Overall, the framework offers a lightweight, data-efficient path to deploy radar-based sensing in real environments with minimal labeling requirements.

Abstract

Learning robust radar perception models directly from real measurements is costly due to the need for controlled experiments, repeated calibration, and extensive annotation. This paper proposes a lightweight simulation-to-real (sim2real) framework that enables reliable Frequency Modulated Continuous Wave (FMCW) radar occupancy detection and people counting using only a physics-informed geometric simulator and a small unlabeled real calibration set. We introduce calibrated domain randomization (CDR) to align the global noise-floor statistics of simulated range-Doppler (RD) maps with those observed in real environments while preserving discriminative micro-Doppler structure. Across real-world evaluations, ResNet18 models trained purely on CDR-adjusted simulation achieve 97 percent accuracy for occupancy detection and 72 percent accuracy for people counting, outperforming ray-tracing baseline simulation and conventional random domain randomization baselines.

Figures (9)

• Figure 1: Simulation setup showing the created radar scenes. A 3D Mixamo humanoid model is animated to represent walking motion and imported into the virtual corridor environment. The FMCW radar observes different scenarios—(left) a single walking subject and (right) two subjects walking at different speeds, producing corresponding range–Doppler signatures.
• Figure 2: Experiment setup for 2 people walking scenario.
• Figure 3: Overview of the proposed sim-to-real pipeline. Yellow blocks: Start of simulated data and real data preprocessing Green path: simulated radar frames from the digital twin are converted to range–Doppler (RD) maps and passed through the noise-floor CDR block, then clipped, normalized, and colour-mapped for network training. Purple path: CDR uses 10 seconds of unlabeled real empty-room data to estimate the target noise-floor statistics that drive the calibration. Blue path: real radar frames for evaluation are converted to RD maps and only clipped and normalized, without any randomization or CDR modification.
• Figure 4: Effect of CDR on global RD magnitude statistics. Baseline simulation (blue) has quite a different low noise floor. CDR (green) shifts the simulated histogram toward the real empty-room distribution (orange).
• Figure 5: Range-Doppler maps for (a,d) baseline simulation, (b,e) simulation after CDR, and (c,f) real measurements for empty room and one-person walking.