Laboratory Near-Field R-ISAR Imaging

• Laboratory near-field R-ISAR is a radar imaging technique that uses controlled rotation and spherical wavefront modeling to capture detailed images of nearby targets.
• It employs both UWB and mmWave modalities with methods like backprojection and analysis-through-synthesis to correct for near-field distortions.
• Advances in neural inverse representations and precise calibration enhance resolution and artifact suppression, improving imaging performance.

Laboratory near-field rotational inverse synthetic aperture radar (R-ISAR) systems deploy controlled rotation or synthetic aperture scan geometries to recover high-resolution, coherent radar images of targets within the near-field zone. Unlike far-field ISAR, where the Fraunhofer distance is large and targets may be considered as point sources, laboratory near-field R-ISAR directly models spherical wavefront propagation, radar cross section (RCS) anisotropy, and geometric distortions. Contemporary advances leverage either traditional backprojection (BP) and holographic reconstruction or modern analysis-through-synthesis (ATS) frameworks combining physical forward models with neural inverse representations. These systems are intrinsically multidisciplinary, involving electromagnetic propagation, array signal processing, inverse problems, and deep learning, and are deployed for purposes ranging from nondestructive testing to robotics and small object recognition (Oshim et al., 2024, Smith et al., 2023).

1. System Architectures and Measurement Configurations

Laboratory near-field R-ISAR systems can be categorized by their radar modalities, aperture synthesis geometries, acquisition protocols, and frequency bands.

Key Modalities

• Ultra-Wideband (UWB) Monostatic ISAR: Employs a wideband pulsed radar (e.g., TimeDomain P440, 3.1–4.8 GHz, center 4.3 GHz, ≈1.7 GHz bandwidth), with both transmitter and receiver co-located. The target is mounted on a motorized turntable to provide cross-range diversity (Oshim et al., 2024).
• Millimeter-Wave (mmWave) MIMO ISAR: Implements MIMO frequency-modulated continuous-wave (FMCW) radar at high frequencies (e.g., Texas Instruments IWR1443-Boost, 77 GHz, 4 GHz bandwidth), using both rotational (R-ISAR) and vertical/planar scanning to synthesize a high-dimensional aperture (Smith et al., 2023).

Setup Elements

Feature	UWB R-ISAR (Oshim et al., 2024)	mmWave MIMO R-ISAR (Smith et al., 2023)
Frequency Band	3.1–4.8 GHz (UWB impulse)	76–80 GHz (mmWave FMCW)
Radar Platform	Monostatic, impulse	2 TX, 4 RX MIMO FMCW
Acquisition	Motorized turntable, 1–360° (1° step)	Rotational (Δθ=0.036°), vertical linear scanning
Range Resolution	~15 cm	~3.8 cm (ΔR = c/(2B))
Cross-Range Control	Rotation of target	Rotation + vertical linear stage (N_y=64, Δy=7.8 mm)
Calibration Objects	Corner reflector, metal sphere	Metal sphere (radius ≈5 mm), corner reflector
Data Rates	350–500 bins/view, 12–360 views	1024 samples/chirp, 360°/64 vertical views
Typical Targets	Soda cans, metal box, NLOS targets	Knife blades, engineered test objects

The systems use absorbers or anechoic panels to suppress multipath. Calibration aligns angle indexing, range gates, and background subtraction, using known reflectors for validation.

2. Near-Field Propagation and Signal Forward Models

The physical forward models in laboratory near-field R-ISAR explicitly accommodate spherical wavefront propagation, near-field amplitude decay, and angle- and aspect-dependent scattering.

UWB R-ISAR Forward Model

• Transmit Pulse:

$$s_t(t) = \exp\left(-\frac{t^2}{2\tau_0^2}\right)\cos(2\pi f_c t)$$

where $f_c=4.3$ GHz, $\tau_0\sim1/\mathrm{BW}$ for bandwidth $\sim1.7$ GHz.

• Received Echo (Point-Scatterer):

$$r(t) = \int_\chi \frac{2\,T(o_T, x)\,b_T(x)}{(4\pi)^3 R^4} L(\sigma(x)) s_t(t-2R/c) dx$$

For range bin at delay $t=2R/c$:

$$r(2R/c) = \int_{E_R} 2\,T(o_T, x)\,b_T(x)\,L(\sigma(x)) dx$$

• Scatter Amplitude (Lambertian Facet):

$$L(\sigma(x)) = \sigma(x)\frac{(x-o_T)}{||x-o_T||} \cdot n(x)$$

• Occlusion (NeRF-inspired):

$$T(o, x_{T_i}) = \prod_{k<i} \exp(-|\sigma_k||l_{k+1}-l_k|)$$

MIMO mmWave R-ISAR Forward Model

• Virtual Aperture: 2 TX × 4 RX yield 8 channels, with vertical displacement $y'$ scanned in discrete steps and a rotary stage covering azimuth $\theta$.
• Multistatic-to-Monostatic Correction:

$$\hat{s}(\theta, k, y') = s(\theta, k, y'_T, y'_R) \cdot \exp\left[-j k \frac{(y'_T - y'_R)^2}{4R_0}\right]$$

Reduces phase errors due to TX/RX separation for $d_y \ll R_0$.

• Range to TX/RX:

$$R_T(\theta, y'_T) = \sqrt{(x - R_0 \cos\theta)^2 + (z - R_0 \sin\theta)^2 + (y - y'_T)^2}$$

A spherical 3D geometry is thus preserved, and all wave curvature and amplitude decay ($1/(R_T R_R)$) is retained.

3. Reconstruction Algorithms: Backprojection and Analysis-Through-Synthesis

Conventional Backprojection (BP)

BP reconstructs the scene reflectivity via time-domain or frequency-domain summation, correcting for near-field wavefront curvature:

$$I_{BP}(x) = \sum_{i=1}^N w(\phi_i) r_i(\tau=2\|x_i - x\|/c) \exp[-j4\pi f_c \|x_i - x\|/c]$$

For 3D mmWave MIMO data, the pipeline involves:

• 2D FFT over (azimuth, vertical)
• Azimuth spatial filtering $g(\theta, k_r) = \exp[j k_r R_0 \cos\theta]$
• Deconvolution, IFFT, and Stolt interpolation to uniform wavenumber grid:

$$k_x = k_r \cos\theta, \quad k_z = k_r \sin\theta, \quad k = \frac{1}{2}\sqrt{k_x^2 + k_y^2 + k_z^2}$$

• Final 3D IFFT yields $\sigma(x, y, z)$.

Analysis-Through-Synthesis (ATS) with Neural Radiance Fields

The ATS approach treats scene reconstruction as a differentiable inverse rendering problem, inspired by NeRF:

• Neural Scene: $F_\theta: x \rightarrow \sigma(x) \in \mathbb{C}$ (complex amplitude), parameterized by a 4-layer MLP (256 units/layer).
• Positional Encoding: Multi-resolution hash encoding (Instant-NGP, 16 levels).
• Sampling: For each view $\phi_i$ and each range shell $R_n$, sample points $x_j$ on the sphere.
• Rendering: Compute synthetic sinogram $\hat{Y}(\theta_{BP})$ by integrating along shells, including occlusion factors and Lambertian reflection.
• Loss Function:

$$\mathcal{L}_{data} = \sum_{i, n} \lVert \hat{r}_i[n] - Y_{i, n} \rVert^2, \quad \mathcal{L}_{reg} = \lVert \theta_{BP} \rVert_2^2$$

$$\mathcal{L} = \mathcal{L}_{data} + \lambda \mathcal{L}_{reg}$$

• Optimization: Minimize $\mathcal{L}$ via Adam, typically for $50k$–$100k$ steps.

The learned $\sigma(x)$ can be backprojected or rendered directly.

4. Experimental Results and Performance Evaluation

UWB ATS vs. BP (Synthetic and Real Data)

• Synthetic Tests: For scenarios with 1–4 point scatterers and additive $CN(0,0.1)$ noise:
  • BP PSNR (1 scatterer): $11.8$ dB
  • ATS PSNR (1 scatterer): $15.0$ dB
  • Similar $3$–$4$ dB gains for multiple scatterers.
  • Under sparsified views (as few as $12$ angles), ATS maintains $2$–$3$ dB higher PSNR.
• Real Measurement: Soda cans, metal boxes, NLOS objects.
  • ATS suppresses side-lobes and artifacts evident in BP (notably under skip-angle/sparse rotation).
  • Substantial improvement in MSE, SNR ($>3$ dB visually, not tabulated).
  • ATS robustly recovers both tags in partial rotation (≤180°), BP fails to resolve multiple targets.
  • For tagged targets (octahedral corner reflectors), ATS localizes at all major orientations; BP returns noisy, ambiguous blobs.

mmWave MIMO R-ISAR (3D Holography)

• Key Metrics:
  • Range resolution: $\Delta R = c/(2B) \approx 3.8$ cm (B = 4 GHz)
  • Vertical: $\approx 1$ mm (aperture $D_y^S=484.8$ mm)
  • Cross-range: $\approx 0.04$ m (radial PSF), azimuth $\delta_{CR} \approx 0.16$ mm (from $\Delta \theta$)
• Artifact Control: Stolt interpolation, spectral windowing, and multistatic phase compensation are required for high-fidelity reconstruction. Regularization (e.g., Tikhonov) may be applied in $k$-domain.

5. Practical Considerations and System Design Guidelines

Hardware Design

• Radar Selection: UWB impulse radar (≥1 GHz) for ~15 cm resolution; mmWave FMCW for finer scaling.
• Antenna Placement: Co-located TX/RX is preferable; include measured antenna pattern $b_T(x)$ in the forward model if separate.
• Motion Control: Motorized turntable with encoder resolution ≤1° (UWB), high-precision rotary and vertical stages (mmWave MIMO).
• Multipath Suppression: Place absorber or anechoic foam behind antennas.

Calibration and Preprocessing

• Time-of-Flight Alignment: Use mechanical zero-range objects (e.g., corner reflectors or metal spheres).
• Background Subtraction: Acquire empty-table sinogram to remove environmental returns.
• Range Profile Normalization: To unit energy or calibrated RCS.
• Matched Filter: Applied to raw A/D samples if not in complex baseband.

Neural Network Training (for ATS)

• Architecture: 4-layer MLP, 256 complex weights/layer; encode position via multi-resolution hash grids.
• Regularization: $\lambda \approx 1\times10^{-4}$ (L₂ weight decay); batch of 1–8 views per iteration, 512–1024 points/range shell.
• Convergence: 50k–100k iterations typical. Slower than BP (≈4.4 s reconstruction on RTX 3080 Ti vs. 0.3 s for BP).
• Limitations: Not real-time; sensitive to antenna/timing misalignment; extension to full 3D reconstructions requires denser coverage.

6. Limitations, Challenges, and Future Directions

• ATS Runtime and Complexity: While ATS delivers superior spatial resolution and artifact suppression (particularly under sparse/noisy scenarios), it remains significantly costlier than BP in runtime and resources (Oshim et al., 2024).
• 3D Volumetric Extension: Achieving volumetric reconstructions in ATS or BP regimes requires denser multi-angular sampling or multi-static geometries. This imposes further hardware and calibration demands (Smith et al., 2023).
• Model Reliance: Accurate knowledge of antenna patterns, precise time and angle alignment, and robust calibration are fundamental; any systematic errors can result in biased or artifact-ridden reconstructions.
• Regularization in Inverse Model: The choice of regularization (e.g., $\ell_2$, total variation, or explicit sparsity) remains open and affects both noise-sensitivity and resolution.
• Strategies for Acceleration: Possible approaches include subsampling $\phi$ or range shells, MLP distillation, or incorporating motion compensation for scenarios with non-constant rotational velocity in practical applications.

The combination of hardware flexibility, rigorous near-field modeling, and recent advances in differentiable rendering and neural field estimation positions laboratory near-field R-ISAR as a robust enabling technology for cost-effective, high-resolution imaging of small, complex objects in controlled and semi-controlled environments. Its methodological interplay with volumetric SAR, computational holography, and implicit scene models marks it as a testbed for future radar imaging research (Oshim et al., 2024, Smith et al., 2023).