Radar Coincidence Imaging

• Radar Coincidence Imaging (RCI) is a computational radar method that forms images by correlating known transmit–receive field patterns with aggregated echo signals across the region of interest.
• The RIS-enabled focused speckle design in RCI enhances signal-to-noise ratio and reduces clutter by leveraging engineered spatial diversity over a limited measurement set.
• Extensions incorporating micro-Doppler and micro-vibration analysis enable phase-sensitive imaging, offering practical applications in security screening and structural health monitoring.

Searching arXiv for foundational and recent papers on Radar Coincidence Imaging and related ghost/computational imaging.
Radar Coincidence Imaging (RCI) is a radar-imaging paradigm in which images are formed not by scanning a focused beam over every point, but by correlating received echoes with a set of known spatial illumination patterns. In this formulation, each measurement is a weighted sum of the scene reflectivity under a deterministic transmit–receive field pattern, and the image is recovered as an inverse problem rather than by point-by-point raster interrogation. Recent work has developed RCI in two technically distinct but conceptually aligned directions: a reconfigurable intelligent surface (RIS)-enabled computational radar architecture that uses focused speckle patterns over a region of interest (ROI) [2507.07285], and a micro-Doppler coincidence-imaging framework that reconstructs the spatial distribution of micro-vibration modes through first-order field correlation and time–frequency-guided processing [2208.13952].

1. Core measurement principle

In its basic form, RCI illuminates the entire ROI with a sequence of known spatial field patterns (E_i(\mathbf r)), (i=1,\dots,M), and records only a small number of measurements for each pattern. The simplest measurement model is a spatial correlation between the scene reflectivity and the transmit–receive field product,

[
g_i \sim \int_{\text{ROI}} E_{\text{Tx}}^{(i)}(\mathbf r)\, E_{\text{Rx}}^{(i)}(\mathbf r)\, \sigma(\mathbf r)\, d\mathbf r,
]

where (\sigma(\mathbf r)) denotes target reflectivity. The defining “coincidence” operation is therefore the correlation between known spatial weighting functions and the aggregated echo.

A discrete forward model writes the measurement process as

[
\mathbf g_{M\times 1} = \mathbf H^{\text{FWD}}_{M\times P}\, \boldsymbol\sigma_{P\times 1},
\tag{1}
]

with (M) measurements, (P) pixels in the forward scene model, measurement vector (\mathbf g), and reflectivity vector (\boldsymbol\sigma). The sensing matrix is populated by the field products at each pixel location (\vec r_j),

[
h_{\text{fwd}(i,j)} = \vec E_{\text{Tx}}^{\,i}(\vec r_j)\cdot \vec E_{\text{Rx}}^{\,i}(\vec r_j).
\tag{2}
]

This is the canonical RCI model: each row of (\mathbf H^{\text{FWD}}) is a known spatial weighting pattern, and each scalar measurement is an inner product between that pattern and the scene. A natural noisy extension is

[
\mathbf g = \mathbf H^{\text{FWD}} \boldsymbol\sigma + \mathbf n.
]

The essential implication is that every configuration “sees” the entire ROI. In contrast to voxel-by-voxel scanning, information is encoded globally across space, so image formation depends on the diversity and conditioning of the measurement operator rather than on one-to-one spatial addressing.

2. Relation to ghost imaging, computational imaging, and conventional radar

RCI is closely related to computational imaging and to ghost imaging. In the ghost-imaging analogy, a known pattern generator illuminates the scene, a bucket detector records an aggregate response, and image formation proceeds by correlation. The RIS-enabled formulation is explicitly placed in a computational imaging framework in which the scene reflectivity is encoded into a relatively small number (M) of measurements by the spatio-frequency diversity of engineered patterns [2507.07285].

The relationship to conventional radar modes is structurally important. Relative to raster scanning, RCI does not require the number of measurements to be at least equal to the number of unknowns, because it uses overlapping, information-rich patterns rather than one focused beam position per voxel. Relative to spotlight or SAR imaging, spatial diversity in RCI can arise from reconfigurable illumination patterns rather than from platform motion. Relative to MIMO radar, the RIS-enabled system discussed in recent work uses a single transmitter and a single receiver together with multiple “virtual masks” generated by programmable surfaces, rather than a physical transmit–receive array.

A common misconception is that coincidence imaging in radar must follow the second-order intensity-correlation formalism of optical ghost imaging. The micro-vibration literature makes the opposite point: in microwave and radar settings, coherent complex-field measurement enables first-order field correlation, which is critical for phase-sensitive imaging and for micro-vibration reconstruction [2208.13952]. In that sense, RCI is not restricted to reflectivity-only recovery from random masks; it also encompasses complex-field reconstruction and phase-coded dynamic imaging.

3. RIS-enabled architectures and focused speckle pattern design

A recent RIS-enabled implementation specifies a single transmitter and a single receiver collocated at ((x=0, y=-3\,\mathrm{m}, z=3\,\mathrm{m})), together with six Tx-RISs facing the transmitter and six Rx-RISs facing the ROI [2507.07285]. Each RIS is a (20\times 20) array with element spacing (d = 2\text{ cm} \approx \lambda/2.5) at 6 GHz and overall size (L_x=L_y=40\text{ cm}). The ROI is located in front of the RISs, around (z=8) m in the main simulation. All RIS elements are assumed ideal, with continuous (0)–(2\pi) phase tuning, no loss, and no mutual coupling. The physical channel is

[
\text{Tx} \rightarrow \text{Tx-RISs} \rightarrow \text{ROI/targets} \rightarrow \text{Rx-RISs} \rightarrow \text{Rx}.
]

Beam redirection at an RIS is modeled by compensating the incident spherical-wave phase and then superimposing a steering gradient,

[
\phi_{d,mn} = k_0 \left( x_{mn} \sin\theta \cos\phi + y_{mn} \sin\theta \sin\phi \right),
\tag{3}
]

where (k_0 = 2\pi/\lambda), ((x_{mn}, y_{mn})) are element coordinates, and ((\theta,\phi)) defines the steering direction. A mask is defined as a complete set of steering angles for all RIS panels, together with per-RIS phase offsets.

The central design feature is the generation of focused speckle: speckle-like interference patterns confined to the ROI and formed by directive beams rather than by fully random element-level fields. Three diversity mechanisms are used. First, per-RIS random phase offsets modify the interference among RISs while preserving the main steering directions. Second, randomized steering angles within ROI bounds introduce small angular deviations while keeping the beams inside the ROI,

[
\theta_{\text{rnd}} = C_\theta (\theta_{\text{max}} - \theta_{\text{min}}), \qquad
\phi_{\text{rnd}} = C_\phi (\phi_{\text{max}} - \phi_{\text{min}}),
\tag{4}
]

with (C_\theta, C_\phi \in [0,0.25]). Third, frequency diversity modifies the propagation phase and induces pattern squinting. In representative simulations, 20 masks and 5 frequency points between 5.9 and 6.1 GHz produce (M=100) measurements; clutter experiments use 30 masks.

This pattern design aims to combine the high signal concentration of directive beamforming with the information richness of computational imaging. The stated consequence is higher signal-to-noise ratio (SNR) and reduced clutter relative to random-pattern metasurface imaging, because both transmit and receive energy are concentrated on the ROI rather than being spread across the environment.

4. Reconstruction, conditioning, and comparison with alternative imaging modes

The inverse problem is formulated by constructing an inverse sensing matrix (\mathbf H^{\text{INV}}_{M\times N}) over a reconstruction grid of (N) voxels, using the same field-product principle as in the forward model,

[
h_{\text{inv}(i,j)} = \vec E_{\text{Tx}}^{\,i}(\vec r_j)\cdot \vec E_{\text{Rx}}^{\,i}(\vec r_j),
]

so that

[
\mathbf g_{M\times 1} \approx \mathbf H^{\text{INV}}_{M\times N}\, \boldsymbol\sigma_{N\times 1}.
]

Because the system is typically underdetermined, with (M < N), the cited RIS-enabled study uses conjugate gradient squared (CGS) to estimate (\tilde{\boldsymbol\sigma}) [2507.07285]. No explicit (\ell_1) or total-variation regularization is introduced; recovery is attributed to measurement diversity and to the moderate sparsity or low complexity of the scenes considered.

The argument for sub-Nyquist measurement counts is therefore operator-theoretic rather than scan-theoretic. Each measurement mixes the whole ROI, and successful inversion depends on the singular spectrum of the sensing matrix. The paper reports that phase offsets alone produce a singular-value decay that saturates after a number approximately equal to the number of Tx-RIS (\times) Rx-RIS combinations, whereas adding angle perturbations and frequency diversity slows the decay and improves effective rank. A plausible implication is that pattern design, rather than raw measurement count alone, is the decisive factor in practical RCI performance.

Three imaging schemes are directly compared under the same RIS hardware and the same total number of measurements. Raster scanning forms a highly directive beam and steers it sequentially across the ROI; with only 20 masks and 5 frequencies, the resulting image is described as “stretched and aliased,” and the 9 subwavelength scatterers cannot be resolved. Random element-level phase patterns provide high spatial diversity but low per-pixel SNR and poor clutter selectivity; at 20 dB SNR, the reconstruction is “highly noisy.” The proposed focused-speckle RCI accurately localizes the 9 scatterers, avoids the aliasing of raster scanning, and remains “robust to clutter” even when strong out-of-ROI scatterers are introduced in the clutter experiment. The evaluation is qualitative rather than metric-based: no explicit PSNR or SSIM values are reported.

5. Micro-Doppler and micro-vibration extensions

RCI has also been extended from static reflectivity imaging to the reconstruction of spatially distributed micro-vibration modes [2208.13952]. In this setting, micro-vibrations have amplitudes much smaller than the wavelength, so their primary radar signature is phase modulation. For line-of-sight displacement (w(t)) at carrier wavelength (\lambda_c), the round-trip phase and micro-Doppler frequency are modeled as

[
\phi(t) \approx \frac{4\pi}{\lambda_c} w(t),
\qquad
f_{\text{micro-Doppler}}(t) \approx \frac{2}{\lambda_c} v(t),
]

where (v(t)=\dot w(t)). Conventional time–frequency analysis can reveal the vibration frequencies in a spectrogram, but it does not recover the spatial distribution of those vibrations across the target.

The coincidence-imaging formulation addresses that limitation by using coded illumination, a single coherent receiver, and a computed reference field. The image functional is a first-order field correlation,

[
G^{(1)}(x_c) = \left\langle E_c^*(x_c,t)\, i(t) \right\rangle,
]

where (E_c(x_c,t)) is the numerically generated reference field and (i(t)) is the coherently detected receiver output. The target is represented as a complex reflectivity whose phase is modulated by vibration,

[
\widetilde{\mathcal T}(x_o,t) = \mathcal T(x_o)\, \exp!\left[j\,2k_\lambda w(x_o,t)\right],
]

with (k_\lambda = 2\pi/\lambda_c). A virtual compensation function is introduced to remove propagation-induced phase terms, leading in the static case to

[
G^{(1)}(x_c) \propto \widetilde{\mathcal T}(x_c).
]

Two target classes are treated. For discrete targets, each point scatterer is assigned vibration modes expanded in Fourier series. A mode-matched temporal compensation term can then be inserted into the virtual modulation so that the reconstructed image contains only the targets sharing a selected vibration mode. The paper reports that different virtual modulations produce different images, each highlighting a subset of the discrete points associated with one vibration mode.

For continuous sheet targets, the displacement is decomposed into principal vibration modes,

[
w(x_o,t) = \sum_{m=0}^{M} W_m(x_o)\, \eta_m(t),
]

where (W_m(x_o)) is the spatial mode shape and (\eta_m(t)) the temporal vibration function. Direct compensation is insufficient because different pixels have different vibration phases. The proposed alternative is interval sampling. Type 1 sampling uses a global period based on the least common multiple of modal sample counts and produces a sequence of complex images over one global period, each image representing the superposition of all modes at a selected phase. Type 2 sampling selects the period of one particular mode and produces predominantly mode-specific reconstructions, though residual cross-mode artifacts remain under finite sampling. This extension shows that RCI can be used not only to recover where scatterers are located, but also to reconstruct where specific dynamic modes are expressed on the target surface.

6. Assumptions, applications, and technical limitations

The current RCI literature represented by these works relies on a set of explicit modeling assumptions. In the RIS-enabled study, the propagation model is narrowband at each frequency point, based on the First-Born approximation, free-space propagation, no mutual coupling, and far-field beamforming geometry [2507.07285]. In the micro-vibration study, assumptions include small vibration amplitude, Fresnel or far-field propagation, perfect knowledge of transmit patterns, coherent stability, periodic vibrations, and idealized or noise-free analysis over many coded realizations [2208.13952]. These assumptions isolate the imaging mechanism, but they also delimit the present empirical scope of the method.

The main application domains identified for RIS-enabled RCI are security screening, wireless user tracking, and activity recognition. The rationale is that multiple RISs can focus and pattern fields over a prescribed ROI using a single Tx/Rx pair and without mechanical scanning. The micro-vibration formulation is positioned as potentially useful for structural health monitoring, mechanical diagnostics, and security or surveillance tasks in which micro-Doppler signatures alone are insufficient because localization of the vibrating substructures is required.

Several limitations recur across the literature. The RIS-enabled simulations assume ideal RIS hardware with continuous phase control, no loss, and no quantization; no detailed robustness analysis is provided for calibration errors, coupling, or control dynamics. The reconstruction methodology is demonstrated numerically and evaluated largely by visual inspection rather than by standardized quantitative image metrics. The micro-vibration work shows residual artifacts from finite sampling and cross-mode leakage, and its interval-sampling strategies require sufficiently separated modal frequencies and adequate sample support. More broadly, both lines of work depend on accurate forward modeling of the illumination patterns, which in practice implies stringent calibration and phase coherence requirements.

Taken together, these developments suggest a broadened view of RCI. It is not only a compressed alternative to raster scanning, nor only a radar analog of optical ghost imaging. It is also a framework for programmable spatial coding, coherent inverse reconstruction, and phase-sensitive imaging of dynamic targets. In that expanded sense, the RIS-enabled and micro-vibration formulations represent two complementary trajectories: one toward electronically reconfigurable, clutter-aware computational radar, and the other toward mode-selective imaging of micro-Doppler phenomena.