Abstract
Conventional imaging systems capture objects visible in the direct line-of-sight (LOS). A decade of research on non-line-of-sight (NLOS) imaging approaches has made it possible to reconstruct hidden geometry outside the line of sight by analyzing indirect light transport. However, most existing methods operate in the optical visible or IR range. Relying on diffuse inter-reflections, every bounce incurs a quadratic intensity falloff. As such, with illumination power limited by eye-safety limitations, existing methods are fundamentally restricted to short ranges on the order of a few meters. We propose an X-band radar-based NLOS imaging method that leverages the long wavelength to convert diffuse reflections into predominantly specular ones, allowing for large-scale hidden-scene perception. We develop a neural reconstruction method that combines a learned dense prediction module and a geometry-aware NLOS reconstruction module, tackling the inherently low spatial resolution of long-wavelength radar. We assess our method with a prototype system and in simulation. Synthetic validation shows that, under the same transmit power, X-band radar achieves 10× longer NLOS reconstruction range than optical systems, while experimental results further demonstrate accurate hidden-object reconstructions up to 40 m, establishing a practical pathway toward real-world long-range NLOS sensing.
Motivation
X-band radar operates at a much longer wavelength than optical and millimeter-wave systems, leading to lower free-space path loss, stronger specular reflection from rough surfaces, and robustness to ambient illumination. The two panels quantify these advantages: free-space path loss (a) and specular reflection ratio (b) across frequency bands show that, at 10 GHz, the signal experiences substantially less attenuation and remains more mirror-like than at 77 GHz or in the optical regime. These propagation and reflection properties make X-band radar particularly suitable for longer-range, large-scale outdoor NLOS imaging.
Radar NLOS Image Formation Model
We propose an X-band radar NLOS image-formation model. A radar wavefront reflects off the relay wall, enters the hidden region, scatters from the object, and returns after a second reflection. In general this is a triple integral over beam patterns, reflectances, attenuation, and path length. But at X-band the three-centimeter wavelength far exceeds the wall roughness, so reflections are specular-dominant. The integral collapses to a single mirror path: the wall acts as a mirror, placing the hidden object in a virtual scene behind it. Steering the radar array then yields the measured range–azimuth (RA) map, which is fed into the reconstruction pipeline.
Method
X-band radar's relatively long wavelength inherently limits its angular resolution. Our neural reconstruction method addresses this by first locating each target as a dense prediction and then mapping every detection back to its true hidden position via a geometry-aware NLOS reconstruction module.
Stage 1: Dense Prediction
A Swin-UNet transformer encoder-decoder backbone captures both local geometric detail and global scene context, producing pixel-wise confidence maps that classify each location as an LOS or mirrored NLOS (mNLOS) response.
Stage 2: Geometry-Aware NLOS Reconstruction
The geometry-aware NLOS reconstruction module turns the dense detections from Stage 1 into real-world hidden positions. A global geometric step reflects each mirrored point across the relay wall, while a local MLP with an attention layer predicts a residual correction. Combined, the analytic reflection plus the learned residual give each hidden point's true position.
Evaluation
Experimental Prototype
We build an X-band radar setup to validate our algorithm. The prototype is an FMCW radar system built on an FPGA (Xilinx ZCU102), operating at 10 GHz with a sweep bandwidth of 700 MHz, yielding a range resolution of approximately 0.21 m. Beam steering is performed by a beamforming front-end driving a 4×8 patch antenna array, scanning from −60° to 60° in 5° increments.
The X-band radar is integrated with an Ouster OS1 LiDAR and a camera, enabling simultaneous capture with LiDAR-derived ground truth and a camera view of the LOS object.
Simulation
We develop an end-to-end radar simulator to generate training data and validate the proposed method. The simulator is grounded in the NLOS image formation model and explicitly integrates four stages: antenna radiation pattern synthesis, FDTD-based material reflectance estimation, multi-path ray tracing, and coherent RA-map formation. Scene geometry and materials are extracted from Unreal Engine, and FDTD simulations characterize the frequency-dependent specular gain for representative urban wall materials at both 10 GHz and 77 GHz.
Dataset
- Pre-train: We pre-train both modules on a synthetic dataset of 2,160 range–azimuth maps. The dataset spans diverse outdoor environments, including urban streets, parking lots, and residential blocks, with varied hidden-object categories such as cars, trucks, buses, and bicycles.
- Fine-tune & Test: We collect 122 captures across 15 distinct scenes under varying conditions, including outdoor daytime and nighttime environments. These captures cover diverse relay-wall materials, including building facades, concrete walls, and metal containers, and hidden-object categories such as cars, vans, and bicycles.
Experimental Results
We visualize real-world NLOS reconstruction results across parking garage, suburban street, and intersection scenes. Each example shows the camera view, radar RA map, dense mNLOS prediction, and final NLOS reconstruction. Estimated object locations are compared against ground truth, with zoomed insets highlighting the agreement between predicted hidden-object clusters and ground-truth vehicle locations.
We compare our method against NLOS-CFAR, RTN, and Further Than CFAR on both synthetic and real-world data, at the dense-prediction (Stage 1) and final NLOS reconstruction (Stage 2) outputs. Each panel plots F1 (higher ↑ is better) against Chamfer Distance (lower ↓ is better) — upper-left corner ↖ is better. Our method consistently occupies the best regime of every plot, confirming gains in both accuracy and geometric reconstruction quality across simulation and real data.
Quantitative Comparison
Quantitative Reconstruction Results. F1 score (higher is better) and Chamfer distance (lower is better) for each method, evaluated on dense prediction (Stage 1) and final NLOS reconstruction (Stage 2) in both simulation and real-world experiments.
BibTeX
@inproceedings{zhao2026xbandnlos,
title = {X-Band Radar Non-Line-of-Sight Imaging},
author = {Du, Dongyu and Zhao, Mingkun and Yang, Yutong and Scheuble, Dominik
and Huang, Xiaolong and Shao, Zijian and Bijelic, Mario
and Sengupta, Kaushik and Heide, Felix},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
year = {2026}
}