MVLA-GR: A Phase-Free Multipath-Based Geometry Reconstruction Method via Multi-View Likelihood Accumulation for ISAC

For voxel $n$ and observation view $m$, consider the event that at least one detected propagation distance is geometrically consistent with the voxel distance, i.e.,

where $R_{m,i}$ is the $i$-th detected propagation distance at view $m$, $r_{m,n}$ is the geometric distance between voxel $n$ and the $m$-th transceiver position, and $\epsilon_{R}$ is a distance tolerance related to the range resolution.

If voxel $n$ corresponds to a true scattering location, the event $\mathcal{E}_{m,n}$ is more likely to occur. In contrast, for an unoccupied voxel, this event can only occur accidentally with a much smaller probability. Based on this observation, we define the hard-decision support indicator

where $\mathbb{I}(\cdot)$ denotes the indicator function.

Treating the $M$ observation views as independent binary trials, the fraction of views in which the event $\mathcal{E}_{m,n}$ occurs provides a coarse approximation to the posterior probability of voxel occupancy:

where $\tilde{P}_{\mathrm{hard}}(\cdot)$ denotes a surrogate posterior probability rather than an exact one. Under this hard-decision model, voxels supported by a larger fraction of observation views are more likely to be occupied.

The hard-decision rule in (23) is not sufficiently robust in practice. First, the extracted distances $R_{m,i}$ are affected by finite bandwidth, peak detection errors, and measurement noise. Second, under viewpoint variation, the dominant scattering point associated with the same physical structure may migrate along the target surface, especially in specular-dominated scenarios. As a result, a strict binary decision may reject physically meaningful observations that exhibit only a small distance deviation.

To address these issues, we replace the hard-decision indicator by a soft probabilistic kernel derived from the statistics of the distance estimation error.

For each detected multipath component, we assume that its corresponding peak in the power delay profile is sufficiently isolated from adjacent components, and that the measurement SNR is sufficiently high. Under these conditions, the extracted propagation distance can be modeled as

where $R_{m,i}$ is the extracted propagation distance, $r^{\star}_{m,i}$ is the true propagation distance of the $i$-th component, and $\varepsilon_{m,i}$ is the associated distance estimation error. We assume that $\varepsilon_{m,i}$ approximately follows a zero-mean Gaussian distribution, i.e., $\varepsilon_{m,i}\sim\mathcal{N}(0,\sigma_{\varepsilon}^{2})$. This assumption is motivated by the asymptotic normality of maximum-likelihood estimators under additive Gaussian noise [20, 9], and is commonly adopted as a tractable approximation in time-delay estimation problems [18]. In practical multipath scenarios where peaks are not fully isolated, this Gaussian model is still used here as an approximate error model for soft geometric matching.

Under this Gaussian error model, the conditional likelihood of observing distance $R_{m,i}$ given that voxel $n$ is occupied and located at geometric distance $r_{m,n}$ is proportional to

where $\sigma_{R}=\sigma_{\varepsilon}$ is the standard deviation of the range estimation error. Defining the distance mismatch

this conditional likelihood motivates the following truncated Gaussian kernel as the soft replacement for the hard indicator $I_{m,n}$:

where $R_{\mathrm{T}}$ is a truncation radius beyond which the voxel is regarded as geometrically incompatible with the observation. The truncation ensures that clearly inconsistent distance observations do not contribute to the voxel support and also improves computational efficiency.

Different detected multipath components do not provide equally reliable geometric evidence and should therefore not contribute equally to voxel occupancy support. The received power $p_{m,i}$ is adopted as a reliability weight for two main reasons.

First, under matched filtering, the delay estimation error $\varepsilon_{m,i}$ decreases with increasing SNR. Components with higher received power generally correspond to higher effective SNR, and their extracted distances are therefore more reliable. Weighting by power naturally suppresses weak and noisy components whose geometric constraints are less trustworthy.

Second, the observed power also reflects the strength of the corresponding propagation path. Stronger detected components usually carry more informative evidence about the underlying scattering structure, while weaker components are more likely to be unstable or noise-contaminated. Therefore, assigning larger weights to stronger components helps the multi-view accumulation focus on more credible geometric evidence.

Accordingly, to preserve the relative reliability differences among components within the same view while preventing any single view from dominating the overall fusion, the weight for each detected component is defined through view-wise normalization:

The view-wise normalization in (28) implicitly assumes that the observation views are of comparable quality, which is typically valid in the considered multi-view sensing settings, where all observations are collected by the same transceiver in a controlled measurement campaign and therefore share similar noise levels and detection conditions. Under this assumption, the relative powers within each view directly reflect the relative reliability of the detected components.

With the soft kernel and power-based weights defined above, the contribution from the $i$-th detected component at view $m$ to voxel $n$ is

Accumulating all detected components within view $m$, we define the total support of voxel $n$ under the $m$-th observation as

where $K_{m}$ is the number of detected multipath components at view $m$.

The quantity $a_{m,n}$ can be interpreted as a surrogate of the posterior probability of voxel occupancy under the $m$-th observation, i.e.,

where $\tilde{P}_{m}(\cdot)$ denotes a surrogate posterior probability rather than an exact one. In this sense, $a_{m,n}$ is the soft generalization of the hard indicator $I_{m,n}$ in (22).

To aggregate evidence across all observation views, we define the normalized multi-view support score of voxel $n$ as

Accordingly, $s_{n}$ can be interpreted as a tractable approximation to the posterior probability of occupancy for voxel $n$ under the full multi-view observations, i.e.,

A voxel with a larger $s_{n}$ is supported by more observation views and more reliable multipath components, and is therefore more likely to be occupied. The collection $\{s_{n}\}_{n=1}^{N}$ forms the reconstructed occupancy map used in all subsequent processing.

The computational complexity of the proposed accumulation process is dominated by the evaluation of voxel-wise contributions for all detected multipath components across all observation views, resulting in

where $N$ is the number of voxels, $K_{m}$ is the number of detected multipath components at view $m$, and $\bar{K}$ denotes the average number of components per view.

In (27), the parameter $\sigma_{R}$ controls how fast the soft support decays as the distance mismatch $d_{m,i}^{(n)}$ increases. As established above, $\sigma_{R}$ should reflect the effective uncertainty of the extracted propagation distance. Although this uncertainty in general depends on multiple factors, including receiver noise, the dominant scale in multipath-rich scenarios is set by the system range resolution $\Delta R=c/(2B)$, which determines the main-lobe width of the pulse shaping function $p(\tau)$ and thereby the achievable precision of peak detection. Accordingly, $\sigma_{R}$ is parameterized as

where $\kappa$ is a dimensionless proportionality constant controlling the tolerance of the radial kernel. A smaller $\kappa$ yields sharper radial discrimination but weaker robustness to range errors, whereas a larger $\kappa$ increases tolerance at the cost of reduced spatial selectivity. In practice, $\kappa$ should be chosen according to the trade-off between range uncertainty and discrimination ability, and is determined in the experimental section.

The parameter $R_{\mathrm{T}}$ in (27) determines the effective support radius of the radial kernel. Under a purely range-limited model, $R_{\mathrm{T}}$ would naturally be chosen on the same order as $\sigma_{R}$. In practical geometry reconstruction scenarios, however, the dominant scattering point associated with the same physical reflecting region may change with the observation viewpoint, especially under specular-dominated scattering conditions. As illustrated in Fig. 3, when the transceiver moves from $\mathbf{T}_{1}$ to $\mathbf{T}_{2}$, the dominant scattering response may shift from one point on the target surface to another nearby point. Although these responses still originate from the same physical reflecting region, their effective propagation distances may differ due to viewpoint-dependent migration of the dominant scattering location.

If $R_{\mathrm{T}}$ is chosen solely according to the range-resolution scale, valid supports associated with the same physical structure may fail to overlap across different views, weakening the subsequent likelihood accumulation. To preserve cross-view overlap among supports corresponding to the same physical scattering region, the truncation radius should also account for viewpoint-induced scattering migration.

To this end, we introduce a migration scale

where $D$ denotes the maximum spatial extent of the target and $\eta$ is a small dimensionless coefficient. The parameter $\delta_{\mathrm{move}}$ characterizes the typical range variation caused by viewpoint-dependent movement of dominant scattering locations. Its value depends on target geometry, surface properties, and angular sampling density, and should be selected to provide sufficient overlap among valid supports across views without excessively enlarging the kernel support.

Since the effective support radius must cover both the range-uncertainty scale and the migration scale, the truncation radius is chosen as

This ensures that the likelihood kernel remains compact enough to suppress clearly incompatible voxels, while still allowing valid supports from different observation views to overlap and accumulate on the same physical scattering structure.

The accumulated support map $\{s_{n}\}_{n=1}^{N}$ provides a continuous spatial representation of voxel occupancy support. However, before converting this support map into a binary reconstruction, spurious responses caused by accidental overlaps among unrelated scattering structures must be suppressed. Such responses appear as reconstruction artifacts, as illustrated in Fig. 4.

A true scattering voxel is expected to be supported not only by a large accumulated response, but also over a sufficiently wide continuous range of observation directions. By contrast, an artifact voxel is usually produced by accidental overlaps among unrelated observations and tends to be supported only within a limited or fragmented angular region. Therefore, in addition to the accumulated response magnitude, the continuity of voxel support across observation directions provides an important cue for artifact suppression.

To characterize this persistence, we first define a binary support indicator for voxel $n$ at view $m$ as

where $a_{m,n}>0$ holds if and only if at least one detected component at view $m$ satisfies $|d_{m,i}^{(n)}|\leq R_{\mathrm{T}}$, as defined in (27).

Let $\phi_{m}$ denote the observation angle of the $m$-th transceiver position with respect to the target center, and assume that $\{\phi_{m}\}$ are ordered increasingly. For voxel $n$, the support sequence $\{b_{m,n}\}_{m=1}^{M}$ generally forms several continuous supporting intervals along the angular axis. Denote these intervals by $\Omega_{n,\ell}$, $\ell=1,\dots,L_{n}$. For each interval $\Omega_{n,\ell}$, its angular width is defined as

The angular support range of voxel $n$ is then defined as the maximum width among all continuous supporting intervals:

To exploit both the response magnitude and the angular persistence, we adopt a joint thresholding strategy. First, a magnitude threshold is applied to the accumulated voxel responses $\{s_{n}\}$, and only voxels whose response values lie within the top $p\%$ of all voxels are retained. Second, an angular-support threshold $T_{\theta}$ is imposed, and only voxels satisfying

are preserved. The joint thresholding criterion is therefore written as

This criterion has a clear physical interpretation: a voxel is regarded as a reliable scattering location only if it exhibits both a high support score and a sufficiently wide continuous support range across observation directions. As a result, spurious voxels caused by accidental support overlap can be effectively suppressed, while voxels associated with true scattering structures are more likely to be retained.

The final binary reconstruction is then obtained by setting $\hat{z}_{n}=1$ for retained voxels and $\hat{z}_{n}=0$ otherwise, thereby yielding an estimate $\hat{\mathbf{z}}$ of the occupancy vector in (20).

Three canonical targets are considered for parameter determination: a circle, a square, and an equilateral triangle. Their geometric sizes are chosen as follows: the circle has a radius of 4 m, the square has a side length of 8 m, and the equilateral triangle has a height of 6 m. These three targets represent smooth curved boundaries, straight edges, and sharp-corner structures, respectively, which correspond to three fundamental scattering mechanisms relevant to MVLA-GR: continuous scattering migration along curved surfaces, piecewise-constant reflection along flat segments, and edge diffraction at sharp corners. Since most practical targets are composed of these three basic scattering elements, parameters determined on this basis are expected to generalize across diverse target geometries. For the circle and square, the maximum extent $D$ is 8 m. For the equilateral triangle, the maximum extent $D$ is taken as its side length, which is $4\sqrt{3}$ m.

In this experiment, each target is observed from 72 uniformly distributed views over the full $360^{\circ}$ angular range, corresponding to an angular interval of $5^{\circ}$. The SNR is fixed to 20 dB, and the remaining system parameters are kept identical across the three targets.

As described in Section III, the parameter $\kappa$ determines the radial kernel scale through $\sigma_{R}=\kappa\Delta R$, while $\eta$ determines the migration scale through $\delta_{\mathrm{move}}=\eta D$.

To determine a unified parameter pair, a grid search is performed over candidate values of $\kappa$ and $\eta$. For each parameter pair $(\kappa,\eta)$, the proposed MVLA-GR method is applied to reconstruct the three canonical targets. During post-processing, the final binary reconstruction is obtained by the joint thresholding rule introduced in Section III, where the retained top $p\%$ support values and the angular-support threshold $T_{\theta}$ are both scanned over their admissible ranges. For each target, the minimum CD obtained over all tested $(p,T_{\theta})$ combinations is taken as the reconstruction error associated with the current parameter pair $(\kappa,\eta)$.

Let $\mathrm{CD}_{q}(\kappa,\eta)$ denote this minimum CD obtained on the $q$-th target, where

The aggregated performance of a parameter pair is defined as the average CD over the three targets:

The final parameter pair is selected as

and is then fixed for all subsequent simulations and measurements in this section.

Fig. 5 shows the parameter sweep results. Fig. 5(a) presents the averaged CD map defined in (44), while Fig. 5(b)–(d) show the corresponding CD maps for the circle, square, and triangle, respectively. The minimum of the averaged CD map is attained at $\kappa^{\star}=0.5$ and $\eta^{\star}=0.04$.

We first compare the proposed method with the baseline under the optimal threshold setting for each configuration. For the proposed MVLA-GR method, the post-processing parameters, namely the retained top $p\%$ support values and the angular-support threshold $T_{\theta}$, are jointly scanned, and the minimum CD over all tested combinations is used as the final result. For the baseline method, the top-$p\%$ threshold is scanned, and the minimum resulting CD is reported.

Fig. 7(a)–(e) shows the reconstruction error as a function of the angular interval under five representative SNR values. In general, the CD increases as the angular interval becomes larger, since fewer observation views are available and the geometric constraints become weaker. Under most tested conditions, the proposed MVLA-GR method achieves lower CD values than the baseline, and the advantage becomes more evident as the angular sampling becomes sparser.

The SNR sensitivity of the proposed method depends strongly on the angular sampling density. When the angular interval is very small, the proposed method is relatively insensitive to SNR, whereas the influence of SNR becomes increasingly visible as the angular interval grows. This indicates that the effects of view density and observation quality are strongly coupled. When the observation views are sufficiently dense, the multi-view geometric constraints are highly redundant, and the accumulated likelihood remains well concentrated around the true contour even if the front-end peak extraction is moderately degraded by noise. As a result, the reconstruction accuracy of the proposed method varies only slightly with SNR in the dense-view regime. In contrast, when the angular interval becomes larger, the number of available views decreases and the geometric constraints become weaker, so the quality of the extracted peaks plays a more critical role, making the reconstruction performance more sensitive to SNR.

Under extremely dense angular sampling, the baseline can become competitive with the proposed method. This is because, in this regime, the observation views are already highly redundant, and the standard backprojection can directly benefit from dense multi-view accumulation. By contrast, the proposed method adopts a unified parameter setting fixed in Section IV-A rather than being re-optimized for this dense-view regime, which may make this unified setting slightly suboptimal when the angular sampling is extremely dense. Nevertheless, as the angular interval increases, the proposed method exhibits a clearer advantage and maintains better reconstruction accuracy under more practically relevant sparse-view conditions.

The baseline appears less sensitive to SNR than the proposed method, because the two methods are affected by noise in different ways. The baseline directly accumulates PDP samples at the corresponding round-trip delays, so noise is partially averaged out during integration. By contrast, MVLA-GR relies on peak extraction before accumulation, so noise additionally affects the front-end observation generation through missed peaks, false peaks, and peak-location perturbations, leading to a clearer dependence on SNR especially when the angular sampling is not sufficiently dense.

Representative reconstruction results of the proposed MVLA-GR method are shown in Fig. 8. The first row illustrates the effect of angular interval when the SNR is fixed at 20 dB, while the second row shows the effect of SNR when the angular interval is fixed at $5^{\circ}$. These visual results are consistent with the quantitative CD curves: denser angular sampling and higher SNR both lead to clearer and more complete reconstruction of the star-shaped contour.

Although the optimal-threshold results reveal the best achievable reconstruction performance under each configuration, such exhaustive tuning is generally unavailable in practice. We therefore further consider a calibrated manual-threshold setting. Specifically, for each SNR, the retained top $p\%$ support ratio and the angular-support threshold $T_{\theta}$ of the proposed MVLA-GR method are fixed according to the average of their optimal values over different angular intervals at that SNR. For the baseline, the top-$p\%$ threshold is calibrated in the same way. The resulting thresholds are then reused for all angular intervals under the corresponding SNR without further tuning. The calibrated threshold values used for each SNR are summarized in Table I.

Fig. 7(f)–(j) shows the corresponding CD curves under this calibrated manual-threshold setting. Compared with the optimal-threshold results in Fig. 7(a)–(e), both methods exhibit performance degradation, which is expected since the thresholds are no longer individually optimized for each configuration. However, the relative performance trend between the two methods remains informative. In particular, as the SNR increases, the proposed MVLA-GR method shows a clearer advantage over the baseline under most angular intervals. This indicates that, although the fixed thresholds introduce some mismatch, the proposed method is still able to exploit multi-view geometric consistency more effectively when the observation quality is sufficiently good.

Table I also shows that the calibrated thresholds of the proposed MVLA-GR method exhibit a clear SNR dependence: the retained top-$p\%$ ratio increases with SNR because cleaner observations contain fewer artifact responses, and $T_{\theta}$ increases monotonically with SNR because more complete and reliable peaks at higher SNR make true scattering voxels remain continuously supported over a wider angular range. By contrast, the calibrated top-$p\%$ threshold of the baseline remains relatively stable across SNR, which is consistent with its low SNR sensitivity discussed above.

The SNR-dependent advantage of MVLA-GR can be explained from two perspectives. When the SNR is moderate or high, the extracted peaks are sufficiently reliable, and the geometric-consistency mechanism of MVLA-GR can still be effectively exploited even under fixed thresholds, so its higher reconstruction ceiling translates into a clear advantage over the baseline. When the SNR is low, peak extraction becomes more vulnerable to missed peaks, false peaks, and location perturbations, which makes MVLA-GR more sensitive to threshold mismatch. Overall, the comparison across Fig. 7 suggests that MVLA-GR not only achieves a higher best-case performance, but also preserves its advantage under calibrated manual thresholds when the SNR is moderate or high, indicating meaningful robustness in practical thresholding conditions.

The measurement is conducted in an anechoic chamber to suppress external interference and unwanted environmental reflections. The measurement platform follows the setup reported in [25], where its calibration accuracy has been verified using a metal sphere as a reference target. A Volkswagen T-CROSS vehicle is used as the target, whose physical dimensions are approximately $4.22\times 1.76\times 1.60~\mathrm{m}$ (length $\times$ width $\times$ height).

The measurement system consists of a vector network analyzer (VNA), two horn antennas, and the associated RF front-end components. The two antennas are placed in close proximity to emulate a quasi-monostatic sensing configuration. During the experiment, the vehicle is mounted on a turntable, while the antennas remain fixed and point toward the target, as shown in Fig. 9. By rotating the turntable, the target is observed from multiple aspect angles, which is equivalent to collecting multi-view measurements around the target. The main measurement configuration is summarized in Table II.

For each observation angle, the VNA records the complex frequency response over the operating bandwidth. After calibration, the corresponding CIR is obtained by inverse Fourier transform, and the PDP is computed as the squared magnitude of the CIR. Since the measured data already contain practical noise and system imperfections, no additional synthetic noise is introduced in this experiment. Following the same phase-free reconstruction framework used in the simulation part, dominant multipath components are extracted from the measured PDPs and represented by their peak distances and peak powers.

The extracted peaks are then fed into the proposed MVLA-GR method. The model parameters are directly inherited from Section IV-A, namely $\kappa^{\star}=0.5$ and $\eta^{\star}=0.04$. Since the antennas have a finite beamwidth rather than omnidirectional radiation patterns, each distance observation is only projected within the corresponding beam-covered angular sector. This beam constraint is incorporated into the likelihood accumulation process to suppress support outside the illuminated region.

The reconstruction region is discretized with a voxel size of 0.005 m. After likelihood accumulation, the final binary result is obtained using manually selected threshold parameters for visualization. Quantitative CD evaluation is omitted here, since accurate voxel-wise ground-truth occupancy labels are not available for the real vehicle target. The reconstruction result is therefore assessed visually.

The measured reconstruction results of the vehicle target are shown in Fig. 10. Fig. 10(a) presents the continuous support map obtained after multi-view likelihood accumulation, while Fig. 10(b) shows the final binary reconstruction result after thresholding.

It can be seen that the continuous support map already reveals the main horizontal contour of the vehicle, although the support remains spatially diffuse before thresholding. After thresholding, the main contour becomes significantly clearer, and the elongated body structure as well as the overall aspect ratio of the vehicle are well preserved. These results indicate that the proposed likelihood accumulation mechanism remains effective under practical measurement conditions.

Compared with the idealized simulation results, the measured reconstruction is less regular and exhibits local discontinuities and some contour thickening. This is expected because the dominant measured scattering points originate from electromagnetically strong parts such as edges, corners, and metallic structures rather than from the outermost physical boundary, the 3D vehicle geometry is projected onto a horizontal imaging plane, and the measured data inevitably contain residual noise, calibration errors, and other non-ideal scattering effects. The recovered contour should therefore be interpreted as the effective projection of dominant scatterers on the imaging plane rather than the exact physical footprint of the vehicle.

Nevertheless, the reconstructed result still preserves the major geometric characteristics of the vehicle. This provides evidence that the proposed MVLA-GR framework is not limited to idealized ray-tracing data, but can also operate on real measured observations.