betaFeedback

WorksComputer Science

Path-Specific Geometric-Stochastic Channel Model for Scattering from ISAC Targets: A Unified Measurement-Based Characterization of Robots, Humans, and Automobiles at 28 GHz

Abstract

This paper develops a channel model for scattering from a broad class of integrated sensing and communication (ISAC) targets-robots, humans, and automobiles-based on a unified measurement campaign at 28~GHz. Nine targets are characterized from numerous transmitter–receiver configurations per target, spanning full bistatic and aspect angles and thus covering backward to forward scattering. Repeated acquisitions are collected for each configuration to capture target motion and dynamics, yielding nearly 117,000 acquisitions in total. Measurements are collected with a 2~GHz-wide channel sounder equipped with a 256-element phased-array antenna, followed by digital beamforming and super-resolution processing for reliable extraction of multiple paths in delay and 2D angle, with average estimation errors of 0.1~ns and 0.2∘^\circ respectively. From these measurements, we reduce a path-specific channel representation in which the scattering is modeled as a cluster of paths whose center is defined by the bistatic specular geometry; the cluster shape is described stochastically. To ensure physically grounded spatio–temporal consistency, extracted paths are mapped to incident points on the target, each with an associated radar cross section (RCS) that is compatible with the 3GPP TR~38.901 ISAC channel model, and are subsequently mapped back to path properties for channel synthesis. The resulting statistical distributions are well behaved, physically interpretable, and consistent across targets, geometries, and motion. Comparison between measured and synthesized channel realizations demonstrates close agreement in pathloss trends and fluctuation behavior. Together, these results yield a high-fidelity, measurement-derived channel representation suitable for ISAC simulation, target modeling, and system design.

Varshney et al.: Multi-Scatterer Target Models for ISAC

Introduction

sensing and communication (ISAC) systems envisioned for future 6G networks require accurate characterization of how dynamic targets—such as unmanned aerial vehicles (UAVs), robots, humans, and automobiles—influence the radio-frequency (RF) propagation channel. Beyond link-level performance, ISAC applications rely on target-dependent scattering behavior to enable sensing tasks including detection, identification, classification, and tracking [1]. Recent multimodal ISAC frameworks further combine RF sensing with modalities such as cameras to improve beam alignment, tracking, and resource allocation in highly mobile networks [35, 36]. Achieving these capabilities requires not only knowledge of aggregate scattering strength, but also an understanding of how scattering geometry evolves with target motion, orientation, and structure. In particular, sensing performance depends on how target-induced multipath is organized spatially and how this structure evolves over time. This need has motivated extensive measurement-based investigations of target radar cross section (RCS) across a wide range of frequencies, target classes, and experimental setups.

A substantial body of recent work reports measured RCS characteristics of individual targets, including UAVs at millimeter-wave (mmWave) bands [22, 16], at microwave bands [16, 18], and for detection performance [7]; road vehicles and infrastructure from 23 GHz to 27 GHz [17]; automotive targets at 94 GHz [9]; and vehicular ISAC scenarios in which cars and trucks act as both reflectors and blockers [40]. Related investigations focus on human targets, providing human signatures for ISAC radar and breath-activity monitoring [19], wideband data for behavior classification [12], near-field pedestrian RCS simulation and validation at 23–28 GHz [39], and pedestrian-oriented ISAC channel measurements [11]. Additional studies characterize object and vehicle scattering using frequency-modulated continuous-wave (FMCW) radars [10] and examine RCS patterns for automotive radar at 77 GHz [21]. While these efforts deliver valuable empirical insight, they employ different hardware platforms, frequencies, geometries, calibration procedures, and extraction methods, which makes it difficult to unify results or derive a consistent target model applicable across ISAC scenarios.

To generalize measured behavior, several works fit statistical distributions such as Lognormal, Nakagami, or Rician models to RCS data in order to capture inter-aspect variability and large-scale scattering effects. Examples include multi-band UAV RCS statistics [28], statistical characterization of UAV and robot signatures [29], 3GPP-compliant RCS characterization of indoor-factory UAV, robotic-arm, and AGV targets [38], and analyses of angular and frequency dependence motivated by ISAC applications [15]. These approaches improve descriptive understanding of target scattering, but they remain fundamentally scalar in nature, providing distributions of total RCS rather than models tied to resolvable propagation paths. As a result, they do not support channel realizations or sensing algorithms that depend on path-level geometry or spatial consistency.

Other studies move toward more structured abstractions by combining measurements with simplified RCS models or ray-based simulations. Several works analyze RCS frequency dependence to inform ISAC system design [23], propose unified target-level RCS factors for link evaluation [31], or investigate monostatic UAV RCS for ISAC channel development [16]. Hybrid approaches combine measurements with raytracing to tune material parameters or validate aggregate scattering behavior [20, 30]. While these models introduce useful structure, validation is typically performed on the total received power summed over all paths, rather than on individual scattering contributions. Because most measurement systems lack sufficient delay and angular resolution, the relationship between target geometry, bistatic configuration, and path-specific scattering behavior remains largely unmodeled.

In parallel, ISAC channel modeling has advanced rapidly, with several frameworks extending traditional communication models to incorporate bistatic geometries and target-induced scattering [24, 14]. Recent models introduce segmented or concatenated ISAC channels that decouple background and target components [8]. This separation is also consistent with the 3GPP TR 38.901 ISAC channel model [2], in which target channel(s) and the background channel are generated separately and then combined. Other models propose multi-scattering-point abstractions [13, 26], formulate general ISAC channel structures combining forward and backward scattering [25], and embed target scattering into hybrid geometric–stochastic models [27]. Despite this progress, most existing models rely on assumed target representations rather than measurement-resolved path data. Moreover, multipoint abstractions often collapse to a single equivalent scatterer after antenna integration, limiting their ability to preserve the geometric detail required for target identification and classification. Consequently, validation is typically restricted to aggregate channel metrics, leaving path-specific scattering structure unverified.

This paper addresses these limitations by presenting a unified measurement campaign coupled with a path-specific scattering model for representative ISAC targets, including robots, humans, and automobiles. All targets are measured using the same 28 GHz channel sounder under consistent geometry, calibration, synchronization, and processing, with multiple transmitter (TX)–receiver (RX) configurations per target selected to systematically span full bistatic and aspect angles, thereby covering backward to forward scattering. A 256-element phased-array receiver enables high-resolution digital beamforming and super-resolution path extraction in post-processing, allowing reliable recovery of multiple paths even for compact targets such as UAVs and robotic arms. Measurement data presented in this paper is available at [37]. The measurements reveal that target scattering is naturally organized as clusters of resolvable paths centered around the specular reflection point. From these observations, we derive a path-specific channel model that maps extracted paths to incident points on the target, providing a common geometric basis for spatial–temporal consistency and incident-point-specific RCS characterization. The main contributions of this paper are as follows:

  1. Unified multi-target measurement campaign: A single 28 GHz ISAC channel-sounding platform is used to measure a total of nine targets–robots, humans, and automobiles. Numerous TX–RX configurations per target are selected to span full bistatic and aspect angles, and repeated acquisitions per configuration to capture target motion and dynamics, yielding nearly 117 000 acquisitions in total.

  2. Super-resolution path extraction: A 256-element phased-array RX with digital (switched) beamforming combined with a 2 GHz probing signal and super-resolution processing enables reliable extraction of multiple paths even from compact targets, achieving average estimation errors of approximately 0.10.1 ns in delay and 0.2∘0.2^\circ in angle.

  3. Path-specific geometric–stochastic channel model: A measurement-derived geometric–stochastic model is developed in which target scattering is represented as clusters of paths anchored at the specular reflection point. By mapping extracted paths to incident points on the target, the model couples path geometry with incident-point-specific RCS statistics while ensuring spatial–temporal consistency.

The remainder of this paper is organized as follows. § 2 describes the measurement campaign. § 3 presents the 28 GHz ISAC measurement system, including the path extraction and path-to-incident-point mapping pipeline. § 4 and § 5 then develop the two coupled components of the proposed path-specific channel model. § 4 defines the geometric component, in which target scattering is organized as a cluster of resolvable paths anchored at the specular reflection point, with cluster shape described stochastically. § 5 introduces the RCS component, which follows a 3GPP-compliant formulation and assigns incident-point-specific scattering strengths within the same incident-point representation. § 6 combines these geometric and RCS components to generate example channel realizations and validate the model. § 7 concludes the paper.

Figure 1, panel 1Figure 1, panel 2Figure 1, panel 3Figure 1, panel 4Figure 1, panel 5Figure 1, panel 6Figure 1, panel 7Figure 1, panel 8Figure 1, panel 9Figure 1, panel 10

Figure 1. Measurement Campaign Summary

Measurement Campaign

A unified measurement campaign was conducted on nine targets: a UAV, a robotic arm, an autonomous mobile robot (AMR), three human subjects1, and three automobiles. Measurement data presented in this paper is available at [37]. Photographs2 of the targets are shown in Fig. fig:targets9. The campaign comprised an ensemble of separate, target-specific variants that spanned different geometric sizes and material properties. In this way, target dimensions are represented through the geometric model, while material-dependent scattering effects are represented empirically through the fitted stochastic parameters rather than through explicit material constants. This approach supports a compact geometric–stochastic target-scatter model, rather than a high-fidelity raytracing or electromagnetic-solver solution.

The objective was to characterize scattering from each target as a function of the aspect angle ϕ\phi and bistatic angle β\beta, consistent with the 3GPP TR 38.901 channel model [2], in which target-specific scattering behavior is parameterized geometrically to support ISAC target-channel and RCS modeling. To this end, two bistatic acquisition configurations were employed. For the UAV, robotic arm, AMR, and humans, the RX was held stationary while the TX was repositioned at nine azimuthal locations spaced 22.5∘22.5^\circ apart on a semicircle surrounding the target (the complementary semicircle was not measured, exploiting the target bilateral symmetry). For each TX position, the targets were set in motion to capture dynamic scattering behavior as well as induce controlled variations in ϕ\phi and β\beta. By contrast, the automobiles were kept stationary and the RX was repositioned at eight azimuthal locations spaced 45∘45^\circ apart on a circle surrounding the target. For each RX position, the TX was mounted on a mobile platform and scanned only a semicircle around the target (again exploiting its bilateral symmetry). This configuration enabled precise and repeatable control of ϕ\phi and β\beta, otherwise unachievable by driving the automobiles.

Across all measurements, TX and RX standoff distances to the target were chosen as short as possible to maximize effective spatial resolution, while ensuring that each target lay fully within the field-of-view (FoV) of both arrays. TX/RX heights and standoff distances were selected per target class to ensure consistent illumination and FoV coverage, with distances for classes containing multiple targets determined based on the largest target in the class. The measurement campaign parameters are summarized in Table Figure 1. The remaining target-specific details are listed below.

  • UAV: A multirotor UAV was mounted on a rotator to vary its aspect angle. For mechanical stability, the platform was placed on the ground and inverted on a low stand so that the underside of the UAV could be observed, which is equivalent to viewing the bottom of the platform in flight.3 At each TX azimuthal position, the UAV was rotated about its vertical axis in 5∘5^\circ steps from 0∘0^\circ to 180∘180^\circ. The target reference point used to define TX/RX standoff distances was the geometric midpoint of the UAV.

  • Robotic arm: An industrial robotic arm with six degrees of freedom was mounted at a height aligned with the TX and RX arrays. During each acquisition, the arm executed a repeatable L-shaped trajectory, consisting of a horizontal sweep from lower right to lower left followed by upward motion. The target reference point used to define TX/RX standoff distances was the base of the robotic arm.

  • AMR: An autonomous mobile robot was instrumented to achieve an effective height of 1.49 m, comparable to a humanoid form factor.4 The AMR executed two linear 2 m trajectories relative to the TX: a radial trajectory directed toward the TX and a tangential trajectory perpendicular to the TX line-of-sight. The target reference point used to define TX/RX standoff distances was the midpoint of each 2 m trajectory.

  • Humans: Three adult subjects were measured: one female (1.60 m) and two males (1.69 m and 1.85 m). Each subject executed a straight 2 m walking trajectory towards the TX, similar to the AMR experiments. The target reference point used to define TX/RX standoff distances was the midpoint of the 2 m walking trajectory.

  • Automobiles: Three automobiles were measured: a sedan, a midsize SUV, and a high-roof van. As explained earlier, the automobiles remained stationary throughout the measurements and the TX was mobile instead. The target reference point used to define TX/RX standoff distances was the midpoint of the automobile body.

Measurement System

This section describes the 28 GHz bistatic channel sounder used in this study. The first subsection describes the RF hardware, the second details the signal processing used to extract discrete propagation paths from the measurements, and third describes the mapping from the extracted paths to their incident points and associated RCSs for downstream modeling.

28 GHz Channel Sounder

Figure 2

Figure 2. Untethered TX and RX of NIST’s 28 GHz bistatic channel sounder. The inset shows the full RX aperture formed by four stacked 8×88\times 8 phased-array boards, including its overall dimensions.

Figure 3
Figure 3
Figure 3

Figure 3. Geometric framework of channel model. (a) Mapping of a measured path to its incident point on the target surface. (b) Cluster center (CC), defined as the intersection of the bistatic specular axis with the canonical ellipsoid representing the target. (c) Lateral perspective of (b), illustrating the conical distribution of incident points around the bistatic specular axis originating at the cluster center.

The measurement system is a 28 GHz channel sounder (Figure 2) developed at NIST, with complete implementation details described in [5]. The TX and RX operate as independent platforms, enabling flexible bistatic configurations. The TX employs a vertically polarized 8×88\times 8 phased-array board that forms a flat-top beam for target illumination. The RX consists of four identical 8×88\times 8 phased-array boards stacked into a single 256-element aperture for enhanced angular resolution. As shown in the inset of Figure 2, the full stacked RX aperture spans 152.4 mm by 108.5 mm. After coherent combination of the four boards, the RX achieves 3.7∘3.7^\circ azimuth and 5.6∘5.6^\circ elevation beamwidths. Because the boards share a common boresight, the effective RX aperture retains the same usable FoV as a single board: an azimuth FoV of −45∘-45^\circ to 45∘45^\circ and an elevation FoV of −25∘-25^\circ to 25∘25^\circ.

The channel is probed using a pseudorandom noise (PN) sequence with 2 GHz bandwidth, centered at precisely 28.5 GHz and transmitted at 30 dBm EIRP. The received signal is downconverted to an intermediate frequency and digitized at 16 GHz. Correlation with the known PN sequence yields complex-valued channel impulse responses (CIRs) that retain both amplitude and phase information across all array elements. The CIRs are calibrated following [4], achieving delay-domain sidelobe suppression of at least 55 dB. Timing synchronization for triggering and phase coherence across acquisitions are maintained via an optical link.

Path Extraction

The element-level, phase-coherent CIRs obtained at the RX are processed using the switched-beamforming framework described in [4]. In place of analog beamforming, the CIRs from all 256 RX elements are planar beamformed in post-processing using calibrated digital weights. The validity of the planar-wave assumption is quantified in Appendix § 8. The calibrated digital weights correct array non-idealities such as distorted or multi-lobed beams that cannot be compensated in analog beamforming due to finite-resolution phase shifters. The digitally synthesized beams achieve angular-domain sidelobe suppression of at least 18 dB [4]. Because the array elements are switched rather than the beams, steering to any direction incurs no increase in channel sweep time, enabling continuous beam synthesis in azimuth and elevation.

The resulting 3D profile in delay, azimuth AoA, and elevation AoA is processed by the Space-Alternating Generalized Expectation-Maximization (SAGE) super-resolution algorithm to extract a discrete set of propagation paths. Path extraction is therefore performed jointly in delay and 2D angle, rather than from the angular response alone. The delay-domain sidelobe suppression of the calibrated CIRs and the angular-domain sidelobe suppression of the digitally synthesized beams together support the extraction of weak paths within the measurement-supported dynamic range. Each path is characterized by a complex amplitude αk=∣αk∣ejφk\alpha_k = |\alpha_k|e^{j\varphi_k}, a delay τk\tau_k, and an angle-of-arrival (AoA) vector θkRX=(θkRX,A, θkRX,E),\boldsymbol{\theta}^{\mathrm{RX}}_k = (\theta_k^{\mathrm{RX,A}},\,\theta_k^{\mathrm{RX,E}}), defined with respect to the RX phased-array boresight uRX\mathbf{u}^{\mathrm{RX}}. Together with xRX=0\mathbf{x}^{\mathrm{RX}}=\mathbf{0}, this unit vector defines the global coordinate system illustrated in Figure 3. Antenna and RF-chain responses are de-embedded following [4], such that αk\alpha_k reflects propagation effects only. Although the inherent RX beamwidths are 3.7∘3.7^\circ (azimuth) and 5.6∘5.6^\circ (elevation), SAGE improves both delay and angle resolution by more than a factor of five [6, 5]. In the controlled validation experiment summarized in Appendix § 9, this extraction pipeline yielded average estimation errors of 0.10.1 ns in delay and 0.2∘0.2^\circ in AoA.

Path-to–Incident-Point Mapping

In their native form, the measured path properties—complex amplitude ∣αk∣ejφk|\alpha_k|e^{j\varphi_k}, delay τk\tau_k, and AoA θkRX\boldsymbol{\theta}^{\mathrm{RX}}_k—are not directly interpretable in geometric terms. To ensure physically grounded spatio-temporal consistency, the proposed channel model is formulated in terms of incident points rather than paths. Each extracted path is therefore mapped to a point xk\mathbf{x}_k on the target surface and to a scattering strength expressed by an intrinsic bistatic RCS σk\sigma_k. Together with the measured phase φk\varphi_k, these quantities define the scattering primitives used in the subsequent sections.

The mapping is implemented using the bistatic range surface–line construction [3], illustrated in Figure 3(a). For each path kk, the TX, RX, and incident point xk\mathbf{x}_k form a bistatic triangle where the measured delay τk\tau_k constrains the total TX–incident point–RX path length through the bistatic range condition

dkTX+dkRX=c τk,\begin{equation} d_k^{\mathrm{TX}} + d_k^{\mathrm{RX}} = c\,\tau_k , \tag*{(1)} \end{equation}

where cc is the speed of light, and

dkTX=∥xk−xTX∥dkRX=∥xk−xRX∥.\begin{align} d_k^{\mathrm{TX}} &= \|\mathbf{x}_k-\mathbf{x}^{\mathrm{TX}}\| \\ d_k^{\mathrm{RX}} &= \|\mathbf{x}_k-\mathbf{x}^{\mathrm{RX}}\|. \end{align}

Since xRX=0\mathbf{x}^{\mathrm{RX}}=\mathbf{0}, the AoA θkRX\boldsymbol{\theta}^{\mathrm{RX}}_k directly defines the RX-side ray, and the incident point lies along this ray as

xk=dkRX[cos⁡θkRX,A cos⁡θkRX,Esin⁡θkRX,A cos⁡θkRX,Esin⁡θkRX,E].\begin{equation} \mathbf{x}_k = d_k^{\mathrm{RX}} \begin{bmatrix} \cos\theta_k^{\mathrm{RX,A}}\,\cos\theta_k^{\mathrm{RX,E}} \\ \sin\theta_k^{\mathrm{RX,A}}\,\cos\theta_k^{\mathrm{RX,E}} \\ \sin\theta_k^{\mathrm{RX,E}} \end{bmatrix}. \tag*{(2)} \end{equation}

Substituting (2) into eq:tx_distance_def and enforcing (1) yields a scalar equation in dkRXd_k^{\mathrm{RX}}, whose solution determines the incident point xk\mathbf{x}_k.

Complete derivation and solution details for the mapping are provided in [3]. The bistatic range surface–line construction is strictly valid for first-order interactions, which is the case here since the proposed model focuses on first-order scatter from the target. To retain only paths whose mapped points lie within the target bounds, 3D gating is applied. The target was intentionally positioned with sufficient spatial separation from other objects to enable this geometric discrimination. Higher-order paths from the target surroundings are not physically interpreted as true incident points, but they nevertheless map outside the known target volume and are therefore filtered out by the gating.

The incident-point RCS is obtained via the calibrated bistatic radar equation. Absolute scaling is established through the metal-sphere calibration [3]. After calibration, the effective TX power and antenna gains satisfy PTX=GTX=GRX=1P^{\mathrm{TX}} = G^{\mathrm{TX}} = G^{\mathrm{RX}} = 1 in linear scale. Under these conditions, the bistatic radar equation reduces to

∣αk∣2=λ2 σk(4π)3 (dkTX)2 (dkRX)2,\begin{equation} |\alpha_k|^2 = \frac{\lambda^2\,\sigma_k} {(4\pi)^3\,(d_k^{\mathrm{TX}})^2\,(d_k^{\mathrm{RX}})^2}, \tag*{(3)} \end{equation}

where λ=c/fc\lambda=c/f_c denotes the wavelength corresponding to the center frequency fcf_c. Solving for the incident-point RCS yields

σk=(4π)3(dkTX)2(dkRX)2λ2 ∣αk∣2.\begin{equation} \sigma_k = \frac{(4\pi)^3 (d_k^{\mathrm{TX}})^2 (d_k^{\mathrm{RX}})^2}{\lambda^2} \,|\alpha_k|^2 . \tag*{(4)} \end{equation}

The phase φk\varphi_k associated with each path is retained to enable coherent superposition in the computation of the composite RCS in § 5.1, which represents the total scattering strength.

Figure 4

Figure 4. Visualization of eight detected paths for a single acquisition of the van target, mapped to incident points on its surface. Marker color denotes the estimated radar cross section (RCS).

Figure 4 shows a representative snapshot of the eight paths extracted for a single acquisition of the van target, mapped to incident points and projected onto the corresponding camera image via Lidar thanks to our context-aware channel sounder [34]. The spatial dispersion of the points across the van highlights the distributed nature of the scattering mechanisms, with contributions arising from multiple structural features such as the rear door, bumper, windows, and underbody. The strongest path within this set corresponds to the vicinity of the bistatic-specular scattering location and appears as the dominant member of the cluster.

Channel Model: Geometry

This section describes the geometric component of the proposed model, in which target scattering is represented as a cluster of paths anchored at the specular reflection point on the target. This point is introduced in the first subsection as the cluster center (CC), a deterministic bistatic–specular reflection point defined by the TX–target–RX geometry. The cluster shape is then characterized by three stochastic parameters: path count, distance offset, and angle offset. These parameters are introduced in the following subsections, and the resulting model parameters are analyzed in the final subsection.

Cluster Center

Measurements show that, for each acquisition, the strongest path most often originates from the surface location satisfying the bistatic specular condition for the given TX\mathrm{TX}–target–RX\mathrm{RX} geometry. For convex targets at mmWave frequencies, this specular contribution dominates and serves as a natural anchor, while nearby diffuse scattering forms a compact cluster around it [32, 33]. As illustrated in Figure 3(b), this location is associated with the surface normal that bisects the incident and reflected rays.

To define this location uniformly across heterogeneous targets, each target is represented by a canonical ellipsoid whose principal axes match its measured height, length, and depth. This ellipsoid is used as a reduced-order geometric reference rather than as a detailed surface model of the target. Although it inevitably simplifies non-convex, articulated, or highly irregular targets, this choice is consistent with the objective of a geometric–stochastic channel model, which seeks a compact parametric representation rather than a high-fidelity raytracing or electromagnetic-solver solution. Deviations between the actual target shape and the canonical ellipsoid are captured by the stochastic component of the model defined in the following subsections.

For a given TX and RX placement, the intersection of the bistatic specular axis with the ellipsoid defines a unique surface point xCC\mathbf{x}^{\mathrm{CC}}, which is referred to as the cluster center (CC). Next, let uTXCC=xTX−xCC\mathbf{u}^{\mathrm{CC}}_{\mathrm{TX}} = \mathbf{x}^{\mathrm{TX}}-\mathbf{x}^{\mathrm{CC}} and uRXCC=xRX−xCC\mathbf{u}^{\mathrm{CC}}_{\mathrm{RX}} = \mathbf{x}^{\mathrm{RX}}-\mathbf{x}^{\mathrm{CC}} denote the vectors from the cluster center toward the TX and RX, respectively. The bistatic specular axis is defined as the bisector of these two vectors,

uspecCC=uTXCC∥uTXCC∥+uRXCC∥uRXCC∥.\begin{equation} \mathbf{u}^{\mathrm{CC}}_{\mathrm{spec}} = \frac{\mathbf{u}^{\mathrm{CC}}_{\mathrm{TX}}} {\|\mathbf{u}^{\mathrm{CC}}_{\mathrm{TX}}\|} + \frac{\mathbf{u}^{\mathrm{CC}}_{\mathrm{RX}}} {\|\mathbf{u}^{\mathrm{CC}}_{\mathrm{RX}}\|}. \tag*{(5)} \end{equation}

This axis acts as the geometric backbone of the cluster, as depicted in Figure 3(b).

Path Count

The number of scattered paths around the cluster center varies stochastically across acquisitions. For each target, the per-acquisition path counts across all measurements were aggregated into an empirical distribution and fit with the Poisson model. This conclusion is supported by Kolmogorov–Smirnov (KS) goodness-of-fit testing across all targets:

f(N)=λNe−λN!,\begin{equation} f(N) = \frac{\lambda^{N} e^{-\lambda}}{N!}, \tag*{(6)} \end{equation}

where the mean λ\lambda is target dependent (Table Figure 5). Representative empirical distributions and Poisson fits appear in Fig. fig:PathCount.

The same aggregation and fitting procedure, together with the KS test against candidate distributions, is used for the distance offset described next and for all subsequent statistical characterizations in this paper.

Distance Offset

For each path kk, let the distance offset

dkCC=∥xk−xCC∥\begin{equation} d_k^{\mathrm{CC}} = \bigl\|\mathbf{x}_k - \mathbf{x}^{\mathrm{CC}}\bigr\| \tag*{(7)} \end{equation}

denote the radial deviation of incident point xk\mathbf{x}_k from the cluster center xCC\mathbf{x}^\mathrm{CC}, as illustrated in Figure 3(c). Across all targets, dkCCd_k^{\mathrm{CC}} was found to obey the Gamma distribution

f(dCC)=(dCC) a−1e−dCC/bba Γ(a),\begin{equation} f(d^\mathrm{CC}) = \frac{(d^\mathrm{CC})^{\,a-1} e^{-d^\mathrm{CC}/b}}{b^{a}\,\Gamma(a)}, \tag*{(8)} \end{equation}

with target-dependent parameters (a,b)(a,b) listed in Table Figure 5. Representative fits are shown in Fig. fig:Distance offset.

Angle Offset

For each path kk, let ukCC=xk−xCC\mathbf{u}^{\mathrm{CC}}_k=\mathbf{x}_k-\mathbf{x}^{\mathrm{CC}} denote the vector from the cluster center toward its reconstructed incident point. The orientation of this vector relative to the bistatic specular axis uspecCC\mathbf{u}^{\mathrm{CC}}_{\mathrm{spec}} is parameterized by the 2D angle offset

θkCC=(θkCC,A, θkCC,E).\begin{equation} \boldsymbol{\theta}^{\mathrm{CC}}_k = \bigl(\theta_k^{\mathrm{CC,A}},\,\theta_k^{\mathrm{CC,E}}\bigr). \tag*{(9)} \end{equation}

The elevation-like deviation from the specular axis is defined as

θkCC,E=cos⁡−1 ⁣(uspecCC⋅ukCC∥uspecCC∥ ∥ukCC∥),\begin{equation} \theta_k^{\mathrm{CC,E}} = \cos^{-1}\!\left( \frac{ \mathbf{u}^{\mathrm{CC}}_{\mathrm{spec}} \cdot \mathbf{u}^{\mathrm{CC}}_k }{ \|\mathbf{u}^{\mathrm{CC}}_{\mathrm{spec}}\|\, \|\mathbf{u}^{\mathrm{CC}}_k\| } \right), \tag*{(10)} \end{equation}

where θkCC,E=0∘\theta_k^{\mathrm{CC,E}}=0^\circ corresponds to perfect alignment with the specular axis and larger values indicate increasing angular deviation.

Across all targets, the empirical distribution of θkCC,E\theta_k^{\mathrm{CC,E}} is strongly forward-weighted, with maximum density near 0∘0^\circ and monotonic decay with increasing angle. This behavior is well modeled by

f(θCC,E)=λ exp⁡ ⁣(−λ θCC,E),\begin{equation} f(\theta^{\mathrm{CC,E}}) = \lambda\,\exp\!\left(-\lambda\,\theta^{\mathrm{CC,E}}\right), \tag*{(11)} \end{equation}

with decay rates summarized in Table Figure 5 and illustrated in Fig. fig:AngleOffset. The azimuthal component θkCC,A\theta_k^{CC,A} specifies rotation about the specular axis and is approximately uniform over [0∘,360∘)[0^\circ,360^\circ), as shown empirically in Fig. fig:AzOffset.

Together with the distance offset dkCCd_k^{\mathrm{CC}}, the elevation offset θkCC,E\theta_k^{\mathrm{CC,E}} defines a conical surface centered on the specular axis. The azimuth offset then selects a unique incident point on this surface. This geometry is depicted in Figure 3(c), which provides a lateral perspective of the configuration shown in Figure 3(b).

Figure 5, panel 1Figure 5, panel 2Figure 5, panel 3Figure 5, panel 4Figure 5, panel 5Figure 5, panel 6Figure 5, panel 7Figure 5, panel 8Figure 5, panel 9Figure 5, panel 10Figure 5, panel 11Figure 5, panel 12Figure 5, panel 13Figure 5, panel 14Figure 5, panel 15Figure 5, panel 16Figure 5, panel 17

Figure 5. Geometry model parameters.

Parameter Analysis

Across the nine measured targets, the fitted parameters in Table Figure 5 and the representative distributions in Figs. fig:PathCount–fig:AngleOffset reveal clear cross-target patterns driven by size, curvature, surface composition, and geometric complexity.

Path count: As established earlier, the number of resolvable paths per acquisition follows a Poisson distribution characterized by a target-dependent rate parameter λ\lambda. For both humans and automobiles, λ\lambda increases monotonically with target size: the largest human and largest automobile produce the most paths, medium-sized subjects fall in the middle, and the smallest yield the fewest. The UAV and robotic arm generate the lowest path counts overall, consistent with their compact dimensions. The AMR falls between these extremes, reflecting its intermediate size and surface extent. A notable exception is the robotic arm, which produces more paths than its physical size alone would suggest. Its metallic, articulated, and multi-faceted structure introduces additional sub-specular scattering mechanisms that elevate λ\lambda beyond what would be expected from geometry alone.

Distance offset: The Gamma-distributed distance offsets dkCCd_k^{\mathrm{CC}} reflect how far resolvable paths extend around the cluster center. The fitted scale parameters follow intuitive geometric patterns: larger targets—particularly the van and SUV, and to a lesser degree the AMR—produce the widest spreads because their extended surfaces support sub-specular returns over a broader region. Human subjects again follow a monotonic size ordering—tallest to shortest— matching the visible torso widths in Figs. fig:targets9(d-f). Compact platforms such as the UAV and robotic arm generate the tightest distributions, consistent with highly localized scattering zones. As with the path-count behavior, the robotic arm shows slightly larger spread than expected for its size due to its complex metallic geometry, which generates multiple nearby reflections around the specular point. These observations demonstrate that the Gamma parameters effectively encode how physical dimensions and surface structure determine the radial width of the scattering cluster.

Angle offset: The angle-offset statistics reveal a physically consistent organization of the scattering cluster across all targets. As established in § 4.4, azimuthal offsets θkCC,A\theta_k^{\mathrm{CC,A}} are approximately uniform about the specular axis, while elevation offsets θkCC,E\theta_k^{\mathrm{CC,E}} are strongly forward-weighted, with maximum density at zero degrees. The emergence of these simple, well-behaved distributions would not occur if the deterministic geometry were misaligned with the underlying scattering mechanisms. Their consistency across disparate target classes therefore confirms that the cluster-centered geometry captures the dominant scattering structure without introducing artificial anisotropy.

The fitted exponential decay rates λ\lambda governing θkCC,E\theta_k^{\mathrm{CC,E}} further quantify the angular concentration of incident points around the specular axis. Since smaller λ\lambda corresponds to slower decay and therefore broader angular spread, the automobiles exhibit relatively broad angular spreads, consistent with their large illuminated surfaces; however, their comparatively smooth body shapes suggest that this spread is driven primarily by physical extent rather than by geometric complexity. By contrast, the human subjects exhibit the largest decay rates and therefore the most concentrated angular spreads, consistent with their smaller size, non-metallic composition, and comparatively smooth body geometry. The robotic targets exhibit broad angular spreads despite their smaller dimensions, which can be attributed to their metallic, articulated, and geometrically complex structures that introduce multiple off-specular scattering contributions. Thus, the angle-offset parameter captures a combined effect of size, material composition, and geometric complexity, rather than target size alone.

Channel Model: RCS

This section describes the RCS component of the proposed model, which assigns scattering strength to the incident-point geometry described in the previous section. It was formulated to remain compatible with 3GPP TR 38.901, the de facto channel model used for wireless system simulation and design. This compatibility allows it to be incorporated directly into 3GPP-style ISAC channel simulations and system-level performance evaluations. Accordingly, the RCS is first modeled at the cluster level through the composite RCS, which represents the total scattered strength across all incident points and follows the 3GPP-style decomposition into mean, deterministic geometry-dependent, and stochastic fluctuation terms. The second subsection introduces a redistribution model that maps the composite RCS to individual incident-point RCS values, thereby coupling the cluster-level RCS characterization to the path-specific geometry. The final subsection interprets the fitted RCS parameters across targets.

Composite RCS

For each acquisition, the composite RCS represents the total coherent scattering strength of the incident points on the target. Given the NN paths along with their corresponding phases φk\varphi_k and incident-point RCS values σk\sigma_k obtained from ((4)), the composite RCS is computed as

σ=∣∑k=1Nσk ej(φk+2πfcτk)∣2,\begin{equation} \sigma = \left| \sum_{k=1}^{N} \sqrt{\sigma_k}\, e^{j(\varphi_k+2\pi f_c\tau_k)} \right|^2 , \tag*{(12)} \end{equation}

where all RCS quantities are expressed in linear units (m2^2) unless stated otherwise. Following 3GPP TR 38.901, the composite RCS is modeled using a multiplicative decomposition as a function of bistatic angle β\beta and aspect angle ϕ\phi,

σ(β,ϕ)=σM σD(β,ϕ) σS,\begin{equation} \sigma(\beta,\phi) = \sigma_{\mathrm{M}}\, \sigma_{\mathrm{D}}(\beta,\phi)\, \sigma_{\mathrm{S}}, \tag*{(13)} \end{equation}

where σM\sigma_{\mathrm{M}} denotes the mean composite RCS, σD(β,ϕ)\sigma_{\mathrm{D}}(\beta,\phi) the deterministic composite RCS capturing angular dependence, and σS\sigma_{\mathrm{S}} the stochastic composite RCS representing small-scale fading. In the sequel, we describe how these three factors are modeled and fitted.

Mean Composite RCS

As defined in (13), σM\sigma_{\mathrm{M}} is the linear mean of the composite RCS across all acquisitions of a given target, thereby averaging over both geometry and small-scale fading variability. It represents the average strength of the cluster. In the 3GPP notation of Table Figure 6, σM\sigma_{\mathrm{M}} is reported in dBsm for each target class.

Deterministic Composite RCS

The deterministic factor σD(β,ϕ)\sigma_{\mathrm{D}}(\beta,\phi) captures the systematic variation of the composite RCS with the bistatic angle β\beta and the aspect angle ϕ\phi. The bistatic angle β\beta describes the relative geometry between the TX, the cluster center (CC), and the RX, and is defined as the angle between the TX–CC and RX–CC vectors,

β=cos⁡−1 ⁣(uTXCC⋅uRXCC∥uTXCC∥ ∥uRXCC∥),\begin{equation} \beta = \cos^{-1}\!\left( \frac{ \mathbf{u}^{\mathrm{CC}}_{\mathrm{TX}} \cdot \mathbf{u}^{\mathrm{CC}}_{\mathrm{RX}} }{ \|\mathbf{u}^{\mathrm{CC}}_{\mathrm{TX}}\|\, \|\mathbf{u}^{\mathrm{CC}}_{\mathrm{RX}}\| } \right), \tag*{(14)} \end{equation}

In contrast, the aspect angle ϕ\phi parameterizes the directional RCS response of the target. It is defined in the RX-centric global coordinate system as

ϕ=cos⁡−1 ⁣(uRXCC⋅uRX∥uRXCC∥).\begin{equation} \phi = \cos^{-1}\!\left( \frac{ \mathbf{u}^{\mathrm{CC}}_{\mathrm{RX}} \cdot \mathbf{u}^{\mathrm{RX}} }{ \|\mathbf{u}^{\mathrm{CC}}_{\mathrm{RX}}\| } \right). \tag*{(15)} \end{equation}

Both angles are illustrated in Figure 3(b).

To isolate the target’s directionality for modeling, each measured composite RCS sweep σ(β,ϕ)\sigma(\beta,\phi) is first circularly shifted by −β/2-\beta/2, corresponding to the bistatic specular axis. This operation aligns the dominant specular response with ϕ=0∘\phi = 0^\circ, yielding a monostatic-equivalent angular response. The aligned sweeps are then combined to obtain a single angular template, σ~D(ϕ)\tilde{\sigma}_{\mathrm{D}}(\phi), modeled using the 3GPP sector function

σ~D(ϕ) (dB)=Gmax⁡−min⁡ ⁣{12 ⁣(ϕ−ϕcenterϕ3dB) ⁣2, σmax⁡},\begin{equation} \tilde{\sigma}_{\mathrm{D}}(\phi)\,(\mathrm{dB}) = G_{\max} - \min\!\left\{ 12\!\left( \frac{\phi - \phi_{\mathrm{center}}}{\phi_{3\mathrm{dB}}} \right)^{\!2}, \,\sigma_{\max} \right\}, \tag*{(16)} \end{equation}

followed by normalization to unit mean in the linear domain. Because the composite RCS also decreases systematically with bistatic angle β\beta, each aligned sweep is normalized by the inverse of a smooth bistatic weighting function

w(β)=k1sin⁡ ⁣(k2β2),\begin{equation} w(\beta) = k_1 \sin\!\left( \frac{k_2 \beta}{2} \right), \tag*{(17)} \end{equation}

which parameterizes the deterministic bistatic dependence.

The deterministic factor is then reconstructed as the product of the angular template and the bistatic weighting,

σD(β,ϕ)=w(β) σ~D ⁣(ϕ+β2),\begin{equation} \sigma_{\mathrm{D}}(\beta,\phi) = w(\beta)\, \tilde{\sigma}_{\mathrm{D}}\!\left(\phi + \frac{\beta}{2}\right), \tag*{(18)} \end{equation}

where σ~D(⋅)\tilde{\sigma}_{\mathrm{D}}(\cdot) is understood in the linear domain. All parameters in (18), including Gmax⁡G_{\max}, ϕcenter\phi_{\mathrm{center}}, ϕ3dB\phi_{3\mathrm{dB}}, σmax⁡\sigma_{\max}, k1k_1, and k2k_2, are jointly optimized from the measurement data.

Finally, since σD(β,ϕ)\sigma_{\mathrm{D}}(\beta,\phi) appears multiplicatively in (13), it is normalized to unit mean, E ⁣[σD(β,ϕ)]=1\mathbb{E}\!\left[\sigma_{\mathrm{D}}(\beta,\phi)\right]=1, to preserve the mean composite RCS. Representative results for σ~D(ϕ)\tilde{\sigma}_{\mathrm{D}}(\phi) appear in Fig. fig:B1RadPattern_UAV, and all fitted deterministic parameters are listed in Table Figure 6.

Figure 6, panel 1Figure 6, panel 2Figure 6, panel 3Figure 6, panel 4Figure 6, panel 5Figure 6, panel 6Figure 6, panel 7Figure 6, panel 8Figure 6, panel 9Figure 6, panel 10

Figure 6. Composite RCS Model Parameters

Stochastic Composite RCS

The stochastic factor captures residual fluctuations in the composite RCS after removing the mean factor σM\sigma_{\mathrm{M}} and the deterministic factor σD(β,ϕ)\sigma_{\mathrm{D}}(\beta,\phi). For each acquisition, the residual is defined as

σS=σ(β,ϕ)σM σD(β,ϕ).\begin{equation} \sigma_{\mathrm{S}} = \frac{\sigma(\beta,\phi)} {\sigma_{\mathrm{M}}\,\sigma_{\mathrm{D}}(\beta,\phi)} . \tag*{(19)} \end{equation}

The stochastic term σS\sigma_{\mathrm{S}} is modeled as a Lognormal random variable with unit mean, E ⁣[σS]=1\mathbb{E}\!\left[\sigma_{\mathrm{S}}\right]=1, to ensure that it does not bias the expected composite RCS under (13). In the log-dB domain, the corresponding distribution is

f ⁣(10log⁡10σS)=1ςσS2πexp⁡ ⁣(−(10log⁡10σS)22 ςσS2),\begin{equation} f\!\left(10\log_{10}\sigma_{\mathrm{S}}\right) = \frac{1}{\varsigma_{\sigma_{\mathrm{S}}}\sqrt{2\pi}} \exp\!\left( -\frac{\left(10\log_{10}\sigma_{\mathrm{S}}\right)^2} {2\,\varsigma_{\sigma_{\mathrm{S}}}^{2}} \right), \tag*{(20)} \end{equation}

where ςσS\varsigma_{\sigma_{\mathrm{S}}} denotes the standard deviation in dB. Representative empirical distributions and fitted models are shown in Fig. fig:B2Dist, and the extracted fluctuation spreads ςσS\varsigma_{\sigma_{\mathrm{S}}} for each target class are listed in Table Figure 6.

Incident-Point RCS

The composite RCS in (13) is formed by the coherent superposition of the incident-point amplitudes σk ejφk\sqrt{\sigma_k}\,e^{j\varphi_k}. Accordingly, the individual RCS values σk\sigma_k represent the power-contributing terms whose coherent sum produces the total cluster strength. To statistically characterize the distribution of point-level RCS values within a cluster, we introduce the normalized quantity

σ~=σσ(β,ϕ),\begin{equation} \tilde{\sigma} = \frac{\sigma}{\sigma(\beta,\phi)}, \tag*{(21)} \end{equation}

where σ(β,ϕ)\sigma(\beta,\phi) denotes the composite RCS modeled in the previous subsection. This normalization removes the absolute cluster strength, isolating the dependence of individual point powers on the underlying scattering geometry.

Measurements show that σ~\tilde{\sigma} decays systematically with the distance offset dCCd^{\mathrm{CC}} from the cluster center. For a given dCCd^{\mathrm{CC}}, the empirical distribution of σ~\tilde{\sigma} is well modeled by a Gaussian distribution in the linear domain,

f ⁣(σ~ ∣ dCC)=1ςσ~ ⁣(dCC)2πexp⁡ ⁣(−(σ~−μσ~ ⁣(dCC))22 ςσ~2 ⁣(dCC)),\begin{equation} f\!\left(\tilde{\sigma} \,\big|\, d^{\mathrm{CC}}\right) = \frac{1}{\varsigma_{\tilde{\sigma}}\!\left(d^{\mathrm{CC}}\right)\sqrt{2\pi}} \exp\!\left( -\frac{ \left(\tilde{\sigma}-\mu_{\tilde{\sigma}}\!\left(d^{\mathrm{CC}}\right)\right)^2 }{ 2\,\varsigma_{\tilde{\sigma}}^{2}\!\left(d^{\mathrm{CC}}\right) } \right), \tag*{(22)} \end{equation}

where both the mean and standard deviation depend on the distance offset from the cluster center. This dependence is parameterized per target class as

μσ~ ⁣(dCC)=aμ ⁣(1−ebμdCC),ςσ~ ⁣(dCC)=aς ⁣(1−ebςdCC)ecςdCC,\begin{align} \mu_{\tilde{\sigma}}\!\left(d^{\mathrm{CC}}\right) &= a_{\mu}\!\left(1 - e^{b_{\mu} d^{\mathrm{CC}}}\right), \\[4pt] \varsigma_{\tilde{\sigma}}\!\left(d^{\mathrm{CC}}\right) &= a_{\varsigma}\!\left(1 - e^{b_{\varsigma} d^{\mathrm{CC}}}\right) e^{c_{\varsigma} d^{\mathrm{CC}}}, \end{align}

where the parameters aμa_{\mu}, bμb_{\mu}, aςa_{\varsigma}, bςb_{\varsigma}, and cςc_{\varsigma} are fitted independently for each target class. These functions capture both the rapid decay of dominant points from the cluster center and the increasing variability of weaker points at larger distance offsets. Representative fits of μσ~(dCC)\mu_{\tilde{\sigma}}(d^{\mathrm{CC}}) and ςσ~(dCC)\varsigma_{\tilde{\sigma}}(d^{\mathrm{CC}}) for the different target classes are shown in Fig. fig:RelativeRCSmodel, and the corresponding parameters are listed in Table Figure 7.

The phases φk\varphi_k were found empirically to be well modeled as independent and uniformly distributed over [0,2π)[0,2\pi), which is expected at 28 GHz: the wavelength is short relative to the separations between resolvable points. This geometric separation decorrelates their phase differences, producing the observed uniform distribution.

Figure 7, panel 1Figure 7, panel 2Figure 7, panel 3Figure 7, panel 4Figure 7, panel 5

Figure 7. Incident-Point RCS Model Parameters

Parameter Analysis

The fitted RCS parameters in Tables Figure 6 and Figure 7, together with the empirical distributions in Figs. fig:B1RadPattern_UAV–fig:RelativeRCSmodel, exhibit clear and physically interpretable trends across the nine measured targets. These trends reflect systematic differences in target size, material composition, facet structure, and geometric complexity, and collectively validate the proposed decomposition of composite and incident-point-level RCS.

Mean Composite RCS: The mean composite RCS σM\sigma_{\mathrm{M}} increases monotonically with overall target size and illuminated surface extent. Human subjects follow a strict ordering with height, while the three automobiles exhibit the expected progression sedan << SUV << van. The AMR and robotic arm occupy intermediate values consistent with their dimensions and metallic construction. For reference, we compare our mean composite RCS values, σM\sigma_{\mathrm{M}}, with published RCS values near 28 GHz for similar target classes, while noting differences in target construction as well as measurement frequency, polarization, and geometry. For UAVs, reported values are −9.6-9.6 dBsm for a mid-size UAV [38] and −9.5-9.5 dBsm for a hexacopter [22], compared with −12.7-12.7 dBsm for our UAV. For robotic platforms, reported values are −8.165-8.165 dBsm for a robotic arm and −11.235-11.235 dBsm for an automated guided vehicle (AGV)-type mobile robot [38], compared with −5.14-5.14 dBsm for our robotic arm and 3.943.94 dBsm for our taller instrumented AMR, respectively. For human targets, prior measurements report a value of −3.78-3.78 dBsm [39], values ranging from −5.47-5.47 dBsm to −2.19-2.19 dBsm across front, rear, and side views [12], and pedestrian extrema spanning −15-15 dBsm to 55 dBsm [40]. These values are consistent in scale with the values of −1.41-1.41 dBsm, 0.760.76 dBsm, and 1.601.60 dBsm obtained for our 1.60 m, 1.69 m, and 1.85 m subjects, respectively. For vehicles, reported sedan extrema range from −15-15 dBsm to 2525 dBsm [40], bracketing the values of 9.989.98 dBsm, 13.3613.36 dBsm, and 20.8820.88 dBsm obtained for our sedan, SUV, and van, respectively.

Deterministic Composite RCS: The deterministic angular dependence σD(β,ϕ)\sigma_{\mathrm{D}}(\beta,\phi), illustrated in Fig. fig:B1RadPattern_UAV and parameterized in Table Figure 6, also follows intuitive structure. Targets with stable, oriented facets—UAV, robotic arm, and all three automobiles—exhibit pronounced directional lobes, whereas humans and the AMR, which lack persistent planar features, show comparatively weak angular variation. The fitted 3GPP sector patterns recover these differences naturally: automobile responses are broad and high-gain, while human and AMR patterns reduce toward the isotropic case. The dependence on bistatic geometry is governed by the scaling parameters k1k_1 and k2k_2, which modulate the angular response as the TX–CC–RX configuration varies.

Stochastic Composite RCS: Examination of the stochastic fluctuation parameter ςσS\varsigma_{\sigma_{\mathrm{S}}} in Fig. fig:B2Dist and Table Figure 6 reveals a consistent cross-target trend. Targets with strong deterministic directionality generally exhibit smaller residual fluctuation variance after removal of the mean and deterministic factors. In these cases, σD(β,ϕ)\sigma_{\mathrm{D}}(\beta,\phi) absorbs much of the geometry-driven variation, leaving less randomness to σS\sigma_{\mathrm{S}}. Human subjects, which are closer to isotropic and exhibit rapidly varying local geometry, show the largest ςσS\varsigma_{\sigma_{\mathrm{S}}} values. The UAV, AMR, and automobiles exhibit lower fluctuation levels, reflecting more stable scattering structure. The van constitutes a notable exception: despite strong directionality, its large physical extent introduces many independent scattering regions whose relative contributions vary across acquisitions, resulting in elevated residual variance. Literature values reported near 28 GHz for ςσS\varsigma_{\sigma_{\mathrm{S}}} include 10.6610.66 dB for a mid-size UAV, 13.5413.54 dB for a robotic arm, and 6.276.27 dB for an AGV [38], compared with 3.743.74 dB, 4.614.61 dB, and 6.776.77 dB for our UAV, robotic arm, and AMR, respectively.

Incident-Point RCS: The RCS redistribution parameters in Table Figure 7 and Fig. fig:RelativeRCSmodel elucidate how the composite RCS is distributed across the incident points. Larger targets such as automobiles and the AMR exhibit slower decay of μσ~(dCC)\mu_{\tilde{\sigma}}(d^{\mathrm{CC}}), consistent with extended illuminated surfaces that support strong near-specular interactions over a wider region. More articulated or irregular targets, including humans and the robotic arm, exhibit larger spreads ςσ~(dCC)\varsigma_{\tilde{\sigma}}(d^{\mathrm{CC}}), reflecting increased variability in how local facets contribute to individual point powers. Compact platforms such as the UAV show both rapid decay and small variance, consistent with a localized scattering region.

Channel Realization

This section describes the algorithm used to generate channel realizations from the reported target-specific channel parameters. The resulting realizations represent the target-channel component of a complete ISAC scene. Environmental clutter is not modeled as part of the target-specific scatter model; instead, it can be generated as a separate background channel and superposed with one or more target channels, consistent with the 3GPP TR 38.901 channel model [2]. The first subsection outlines the target-channel realization procedure, while the second presents representative realizations.

Channel Realization Procedure

The instantaneous channel realization at time tt is constructed by combining the geometric scattering framework of § 4 with the RCS model of § 5. As illustrated in Figure 8, the realization proceeds across three hierarchical levels: 1) the cluster level; 2) the incident-point level; and 3) the path level. The green blocks denote the geometric model parameter inputs, the yellow blocks the stochastic model parameter inputs, and the red blocks the intermediate level inputs/outputs.

1) Cluster Level

As illustrated in Figure 8, the cluster level determines the cluster center and bistatic specular axis and the composite RCS.

Figure 8

Figure 8. Hierarchical channel realization flowchart. 1) Cluster-level realization. 2) Incident-point-level realization. 3) Path-level realization.

Cluster geometry: In the general case, the target is modeled as a canonical ellipsoid at time tt, consistent with § 4.1, which uniquely determines the bistatic–specular cluster center xCC(t)\mathbf{x}^{\mathrm{CC}}(t) and the associated bistatic specular axis uspecCC(t)\mathbf{u}^{CC}_{\mathrm{spec}}(t) from § 4.1. For nondirectional targets, the ellipsoid may be replaced by a sphere and, in the limiting case of a compact target (negligible sphere radius), by a point scatterer.

Composite RCS: Given the instantaneous geometry, the composite RCS σ(t)\sigma(t) is evaluated using the multiplicative model in (13). The required bistatic angle β(t)\beta(t) and aspect angle ϕ(t)\phi(t) follow directly from the instantaneous geometry via (14) and (15), and the deterministic and stochastic factors are applied as described in § 5.1.

2) Incident-Point Level

As illustrated in Figure 8, the incident-point level stochastically instantiates N(t)N(t) incident points around the cluster geometry and redistributes the composite RCS across them. The number of paths N(t)N(t) associated with the target is drawn according to the Poisson distribution in (6) in § 4.2.

Incident-point geometry: For each path k=1,…,N(t)k=1,\dots,N(t), an incident point xk(t)\mathbf{x}_k(t) is generated relative to xCC(t)\mathbf{x}^{\mathrm{CC}}(t) by

  • drawing the distance offset dkCC(t)d_k^{\mathrm{CC}}(t) according to the Gamma distribution in (8) from § 4.3;

  • drawing the elevation angle offset θkCC,E(t)\theta_k^{\mathrm{CC,E}}(t) according to the exponential distribution in (11) from § 4.4;

  • drawing the azimuthal angle offset θkCC,A(t)\theta_k^{\mathrm{CC,A}}(t) uniformly over [0∘,360∘)[0^\circ,360^\circ).

These parameters uniquely determine the 3D incident point xk(t)\mathbf{x}_k(t) of each path.

Incident-point RCS: The composite RCS is redistributed across the N(t)N(t) incident points using the path-level RCS model in § 5.2. For each path, a normalized RCS σ~k(t)\tilde{\sigma}_k(t) is drawn conditionally on dkCC(t)d_k^{\mathrm{CC}}(t) according to (22)–(?), and the absolute path RCS is obtained as σk(t)=σ~k(t) σ(t).\sigma_k(t)=\tilde{\sigma}_k(t)\,\sigma(t).

3) Path Level

As illustrated in Figure 8, each realized incident point and its associated RCS map to a set of path properties, which correspond to the inverse of the path-to–incident-point mapping in § 3.3. These properties are then used to synthesize the instantaneous channel response h(t)h(t).

Path geometry: For each incident point xk(t)\mathbf{x}_k(t), the geometric path parameters follow directly from the TX and RX distances to xk(t)\mathbf{x}_k(t) given by dkTX(t)d_k^{\mathrm{TX}}(t) and dkRX(t)d_k^{\mathrm{RX}}(t) in (?). Their sum defines the total bistatic path length dk(t)=dkTX(t)+dkRX(t)d_k(t)=d_k^{\mathrm{TX}}(t)+d_k^{\mathrm{RX}}(t), which satisfies the bistatic range condition (1) and yields the propagation delay τk(t)=dk(t)/c\tau_k(t)=d_k(t)/c. Under the first-order interaction assumption, the incident point also uniquely determines the AoD at the TX and the AoA at the RX. Let x=[x,y,z]T\mathbf{x}=[x,y,z]^{\mathsf T} denote a generic Cartesian position vector. The AoD is computed as

θkTX,A(t)=tan⁡−1 ⁣(yk(t)−yTX(t)xk(t)−xTX(t))θkTX,E(t)=sin⁡−1 ⁣(zk(t)−zTX(t)∥xk(t)−xTX(t)∥)\begin{aligned} \theta_k^{\mathrm{TX,A}}(t) &= \tan^{-1}\!\left( \dfrac{y_k(t)-y^{\mathrm{TX}}(t)} {x_k(t)-x^{\mathrm{TX}}(t)} \right)\\[4pt] \theta_k^{\mathrm{TX,E}}(t) &= \sin^{-1}\!\left( \dfrac{z_k(t)-z^{\mathrm{TX}}(t)} {\|\mathbf{x}_k(t)-\mathbf{x}^{\mathrm{TX}}(t)\|} \right) \end{aligned}

and the AoA as

θkRX,A(t)=tan⁡−1 ⁣(yk(t)−yRX(t)xk(t)−xRX(t))θkRX,E(t)=sin⁡−1 ⁣(zk(t)−zRX(t)∥xk(t)−xRX(t)∥)\begin{aligned} \theta_k^{\mathrm{RX,A}}(t) &= \tan^{-1}\!\left( \dfrac{y_k(t)-y^{\mathrm{RX}}(t)} {x_k(t)-x^{\mathrm{RX}}(t)} \right)\\[4pt] \theta_k^{\mathrm{RX,E}}(t) &= \sin^{-1}\!\left( \dfrac{z_k(t)-z^{\mathrm{RX}}(t)} {\|\mathbf{x}_k(t)-\mathbf{x}^{\mathrm{RX}}(t)\|} \right) \end{aligned}

Path complex amplitude: Each path RCS σk(t)\sigma_k(t) is then mapped to the magnitude of the complex path coefficient ∣αk(t)∣|\alpha_k(t)| through the bistatic radar equation in (3). An independent random phase φk(t)∼U[0,2π)\varphi_k(t)\sim\mathcal{U}[0,2\pi) is assigned, yielding αk(t)=∣αk(t)∣ejφk(t).\alpha_k(t)=|\alpha_k(t)|e^{j\varphi_k(t)}.

Channel synthesis: Given the complete set of path properties at time tt, the instantaneous channel response is synthesized in the joint angle–delay domain as

Pathloss: The corresponding instantaneous pathloss is defined as

PL(t)=1∣h(t)∣2.\begin{equation} PL(t)=\frac{1}{|h(t)|^2}. \tag*{(23)} \end{equation}
Figure 9, panel a

(a) UAV

Figure 9, panel b

(b) Robotic Arm

Figure 9, panel c

(c) AMR

Figure 9, panel d

(d) Human--1.85\,m Male

Figure 9, panel e

(e) Automobile--Van

Fig.  . Realized versus measured pathloss for five representative targets.

Figure 9.

Example Realizations

A software implementation written in MATLAB to realize the channel model for all targets is available here [41]. Fig. fig:PLcomparison compares the measured and realized pathloss evolution for five representative target classes under dynamic motion. In all cases, both measured and realized pathloss are computed identically using the channel synthesis and pathloss definitions in eq:ht and (23), i.e., by coherently summing all resolved target-associated paths and taking the inverse of the resulting channel power gain. For the first half of the observation window, the realized trajectories exactly replicate the measured motion, where acquisitions are indexed by the time variable tt; this interval therefore serves as a direct validation of the proposed geometric and RCS-based model under observed dynamics. During the second half, controlled trajectory variations are introduced that deviate from the measured motion, for extrapolation beyond the measurement conditions.

For the UAV case, the first half corresponds to a continuous rotation from 0∘^\circ to 180∘^\circ at a fixed location, capturing aspect-dependent RCS variations; the resulting pathloss minima align with measured RCS peaks at favorable aspect angles. In the second half, the UAV is assumed to recede from the TX and RX at constant speed, producing the increasing pathloss trend toward the end of the interval. For the robotic arm, the realized motion follows the measured articulated trajectory during the first half, while in the second half the arm configuration is driven away from the measured trajectory to the extent permitted by the arm reach, producing a gradual pathloss increase under physically feasible articulated motion. For the human and AMR cases, the realized trajectories follow the measured approach motion during the first half, resulting in reduced pathloss due to decreasing bistatic separation and improved geometric alignment. In the second half, both targets move away from the TX and RX, leading to a gradual pathloss increase consistent with increasing propagation distance. For the van case, the TX follows the measured motion initially, after which it remains stationary while the automobile recedes from both TX and RX, producing a monotonic pathloss increase; the pathloss reduction observed while the TX is moving coincides with measured RCS peaks from the automobile side.

Overall, the close agreement in pathloss trends and fluctuation statistics across both validation and extrapolation intervals demonstrates that the proposed target-aware channel modeling framework captures motion-induced pathloss dynamics across diverse target classes.

Conclusion

This paper presented a unified channel model for scattering from representative ISAC targets based on an exhaustive campaign of 28 GHz channel measurements. By combining a 256-element phased-array receiver with super-resolution processing, multiple paths were reliably extracted across all target classes, enabling a path-specific view of target-induced scattering that is directly tied to physical geometry. Target scattering was shown to be well represented by clusters of paths anchored at a bistatic-specular cluster center defined by the TX–target–RX geometry, providing an explicit and interpretable link between propagation geometry and target shape. A deterministic geometric backbone ensures spatial–temporal consistency as targets move and reorient, while a stochastic description tuned directly from measurements captures the observed variability in scattering behavior without sacrificing physical intuition.

Across nine heterogeneous targets, the resulting model parameters exhibit consistent and interpretable trends: larger and more extended targets produce broader clusters, while material composition and geometric complexity introduce systematic variations in angular spread and power fluctuations. The inferred statistical behavior is well behaved and stable across geometries and motion, supporting robust channel realization. Validation through synthesized channel realizations demonstrates close agreement with measured pathloss evolution under dynamic motion, confirming that the proposed model faithfully reproduces both large-scale trends and path-level fluctuations. Together, these results establish a path-specific, geometry-driven, and measurement-tuned channel representation that preserves target shape information while remaining compatible with communication-oriented abstractions. As such, the proposed model is well suited for ISAC network simulation and evaluation.

Here we validate the far-field planar-wave assumption for our RX array. At 28.5 GHz, the Fraunhofer distance 2D2/λ2D^2/\lambda—the conventional far-field threshold corresponding to a maximum center-to-edge phase deviation of 22.5∘22.5^\circ—is 4.4 m for the RX aperture (D=152.4D=152.4 mm). However, the minimum RX–target distance in the measurement campaign is 2.5 m and thus lies within the near-field region, for which the corresponding phase deviation is approximately 40∘40^\circ. To quantify the effect of spherical-wave curvature within the near-field using the actual RX element geometry, spherical wavefronts were generated and processed using the array’s far-field planar-wave beamforming weights. Figure 10 shows the magnitude and phase cuts of the beam pattern at azimuth 0∘0^\circ for several finite scatterer distances, together with the far-field pattern. The complementary cut at elevation 0∘0^\circ was also evaluated and exhibited similar behavior.

The magnitude plots in Figure 10(a) show rapid convergence of the near-field pattern toward the far-field. The pattern at 2.5 m remains close to the 4.4 m Fraunhofer pattern (as well as the far-field pattern), particularly in the main-lobe region, whereas the most pronounced deviations occur below −23-23 dB, far from the main peak. The shaded region in Figure 10(a) highlights the difference between the 2.5 m and 4.4 m patterns. The normalized root-mean-square error (NRMSE) with respect to the far-field pattern, computed in linear magnitude over the full angular cut, is 9.46 % at 2.5 m. The error decreases further to 5.15 % at 4.4 m, mainly due to reduced discrepancies in the low-level sidelobes. Because the magnitude patterns in Figure 10(a) are normalized by the same far-field peak, any array-gain defocusing, quantified as near-field peak-gain reduction, remains visible. For a scatterer at the azimuth 0∘0^\circ steering direction, the gain at 2.5 m is 0.190.19 dB below the far-field value, which is within the 0.48 dB path-gain error reported for the measurement system in [4]. Therefore, the array-gain defocusing causes a negligible estimation error in the scatterer’s path gain or RCS.

The phase plots in Figure 10(b) also show rapid convergence of the near-field pattern toward the far-field reference. However, even at 100 m, more than an order of magnitude beyond the 4.4 m Fraunhofer distance, the near-field pattern does not reach the ideal far-field. This suggests that the Fraunhofer distance should be interpreted as a guideline rather than a sharp boundary. The residual phase difference between 2.5 m and 4.4 m, shaded in Figure 10(b), remains below 7∘7^\circ over the angular region considered here, comparable to the approximately 6∘6^\circ standard deviation of the measurement system’s phase noise reported in [4]. Thus, the phase deviation introduced by using planar-wave beamforming at 2.5 m is on the order of the system phase uncertainty.

Figure 10, panel 1Figure 10, panel 2

Figure 10. Far-field and near-field RX beamforming patterns at several scatterer distances: (a) normalized magnitude beam pattern and (b) referenced beam phase pattern. The shaded region highlights the difference between the patterns at the minimum measurement distance of 2.5 m and the 4.4 m Fraunhofer distance.

The path-parameter extraction accuracy was evaluated using a controlled experiment with known geometry. The receiver was mounted on a rotator and measured at 120 control points, producing known changes in path length and AoA. The extracted delay and AoA were then compared against the corresponding values computed from the known geometry. Figure 11 shows the resulting empirical CDFs. The average errors are 0.10.1 ns in delay and 0.2∘0.2^\circ in AoA.

Figure 11, panel 1Figure 11, panel 2

Figure 11. Empirical CDFs of path-parameter estimation errors from the controlled validation experiment: (a) delay error and (b) AoA error.

References

  1. [1]Wang, J., Varshney, N., Gentile, C., Blandino, S., Chuang, J., & Golmie, N. (2022). Integrated sensing and communication: Enabling techniques, applications, tools and data sets, standardization, and future directions. IEEE Internet of Things Journal, 9(23), 23416–23440.
  2. [2]TR 38.901 V19.0.0: Study on Channel Model for Frequencies from 0.5 to 100 GHz. (2025). TR 38.901 V19.0.0: Study on Channel Model for Frequencies from 0.5 to 100 GHz. In 3rd Generation Partnership Project (3GPP) (Numbers 3GPP TR 38.901). https://www.3gpp.org3gpp.org
  3. [3]Gentile, C., Chuang, J., Blandino, S., Senic, J., Bang, J., & Berweger, S. (2025). Measuring the RCS of Multipoint-Scattering Targets in JCAS Channel Sounding. 2025 19th European Conference on Antennas and Propagation (EuCAP), 1–5. https://doi.org/10.23919/EuCAP63536.2025.10999612
  4. [4]Caudill, D., Chuang, J., Jun, S. Y., Gentile, C., & Golmie, N. (2021). Real-time mmWave channel sounding through switched beamforming with 3-D dual-polarized phased-array antennas. IEEE Transactions on Microwave Theory and Techniques, 69(11), 5021–5032.
  5. [5]Chuang, J., Caromi, R., Senic, J., Berweger, S., Varshney, N., Wang, J., Lai, C., Bodi, A., Sloane, W., Gentile, C., et al. (2024). Quasi-deterministic channel propagation model for human sensing: Gesture recognition use case. IEEE Open Journal of Antennas and Propagation, 5(3), 557–572.
  6. [6]Papazian, P. B., Choi, J.-K., Senic, J., Jeavons, P., Gentile, C., Golmie, N., Sun, R., Novotny, D., & Remley, K. A. (2016). Calibration of millimeter-wave channel sounders for super-resolution multipath component extraction. 2016 10th European Conference on Antennas and Propagation (EuCAP), 1–5.
  7. [7]Rosamilia, M., Balleri, A., De Maio, A., Aubry, A., & Carotenuto, V. (2022). Radar detection performance prediction using measured UAVs RCS data. IEEE Transactions on Aerospace and Electronic Systems, 59(4), 3550–3565.
  8. [8]Jiang, T., Liu, G., Xia, L., Jin, J., Wang, Q., Li, Y., & Zhang, J. (2025). The Segmented Convolutional Channel Model for ISAC. 2025 10th International Conference on Computer and Communication System (ICCCS), 738–743.
  9. [9]Kamel, E. B., Peden, A., & Pajusco, P. (2017). RCS modeling and measurements for automotive radar applications in the W band. 2017 11th European Conference on Antennas and Propagation (EUCAP), 2445–2449.
  10. [10]Lee, S., Kang, S., Kim, S.-C., & Lee, J.-E. (2016). Radar cross section measurement with 77 GHz automotive FMCW radar. 2016 IEEE 27th Annual International Symposium on Personal, Indoor, and Mobile Radio Communications (PIMRC), 1–6.
  11. [11]Ebrahimizadeh, J., Cai, X., Sanchez, J., Vinogradov, E., Tufvesson, F., & Vandenbosch, G. A. (2024). Millimeter-Wave Scattering Model Based on RCS Distribution: A Simulation and Verification Study. 2024 IEEE Globecom Workshops (GC Wkshps), 1–6.
  12. [12]FORTUNY, G. J., CHAREAU, J.-M., et al. (2013). Radar Cross Section Measurements of Pedestrian Dummies and Humans in the 24/77 GHz Frequency Bands.
  13. [13]Yang, Y., Gradoni, G., Elzanaty, A., Tishchenko, A., Khalily, M., Tafazolli, R., Heggo, M., Shojaeifard, A., Mestrah, Y., Hemadeh, I., et al. (2025). ISAC Channel Modelling with Multi-Scattering-Point RCS. 2025 IEEE 26th International Workshop on Signal Processing and Artificial Intelligence for Wireless Communications (SPAWC), 1–5.
  14. [14]Wei, Z., Jia, J., Niu, Y., Wang, L., Wu, H., Yang, H., & Feng, Z. (2024). Integrated sensing and communication channel modeling: A survey. IEEE Internet of Things Journal.
  15. [15]Fenollosa, S., Drozdowska, M., Yang, W., Micó-Rosa, S., Castilla, A., Lopez-Escudero, A., Li, J., & Cardona, N. (2024). Frequency-responsive RCS characteristics and scaling implications for ISAC development. 2024 IEEE Globecom Workshops (GC Wkshps), 1–6.
  16. [16]Yuan, Z., Yu, L., Wang, Z., Li, C., Dallmann, T., & Fan, W. (2024). Experimental analysis and modeling of mono-static aav RCS for ISAC channels. IEEE Antennas and Wireless Propagation Letters.
  17. [17]Schipper, T., Fortuny-Guasch, J., Tarchi, D., Reichardt, L., & Zwick, T. (2011). RCS measurement results for automotive related objects at 23-27 GHz. Proceedings of the 5th European Conference on Antennas and Propagation (EUCAP), 683–686.
  18. [18]Semkin, V., Yin, M., Hu, Y., Mezzavilla, M., & Rangan, S. (2021). Drone detection and classification based on radar cross section signatures. 2020 International Symposium on Antennas and Propagation (ISAP), 223–224.
  19. [19]Piuzzi, E., D’Atanasio, P., Pisa, S., Pittella, E., & Zambotti, A. (2015). Complex radar cross section measurements of the human body for breath-activity monitoring applications. IEEE Transactions on Instrumentation and Measurement, 64(8), 2247–2258.
  20. [20]Luo, C., Tang, A., Gao, F., Liu, J., & Wang, X. (2024). Channel modeling framework for both communications and bistatic sensing under 3GPP standard. IEEE Journal of Selected Areas in Sensors.
  21. [21]He, D., Guan, K., Ai, B., Zhong, Z., Kim, J., Chung, H., & Hrovat, A. (2022). Channel measurement and ray-tracing simulation for 77 GHz automotive radar. IEEE Transactions on Intelligent Transportation Systems, 24(7), 7746–7756.
  22. [22]Semkin, V., Haarla, J., Pairon, T., Slezak, C., Rangan, S., Viikari, V., & Oestges, C. (2020). Analyzing radar cross section signatures of diverse drone models at mmWave frequencies. IEEE Access, 8, 48958–48969.
  23. [23]Zhang, Y., Zhang, J., Gong, H., Hu, X., Zhang, J., Xing, H., Luo, S., Xiong, Y., Yu, L., Yuan, Z., et al. (2025). A unified rcs modeling of typical targets for 3gpp isac channel standardization and experimental analysis. IEEE Journal on Selected Areas in Communications.
  24. [24]Ghosh, A., Wild, T., Du, J., Tan, J., Grudnitsky, A., Chizhik, D., Mandelli, S., Xing, Y., Schaich, F., & Viswanathan, H. (2025). A Unified Future: Integrated Sensing and Communication (ISAC) in 6G. IEEE Journal of Selected Topics in Electromagnetics, Antennas and Propagation.
  25. [25]Yang, R., Wang, C.-X., Huang, J., Aggoune, E.-H. M., & Hao, Y. (2023). A novel 6G ISAC channel model combining forward and backward scattering. IEEE Transactions on Wireless Communications, 22(11), 8050–8065.
  26. [26]Ye, Q., Huang, Y., Luo, Q., Hu, Z., Zhang, Z., Zhao, Q., Su, Y., Hu, S., & Yang, G. (2024). A general integrated sensing and communication channel model combined with scattering clusters. IEEE Transactions on Vehicular Technology.
  27. [27]Zhang, Z., He, R., Ai, B., Yang, M., Li, C., Mi, H., & Zhang, Z. (2022). A general channel model for integrated sensing and communication scenarios. IEEE Communications Magazine, 61(5), 68–74.
  28. [28]Das, S., Zhang, P., Hovinen, V., Leinonen, M. E., Pärssinen, A., & Kyösti, P. (2025). Monostatic and Bistatic RCS Characterisation of a Drone for ISAC Channel Modelling. Authorea Preprints.
  29. [29]Waqar Azim, A., Bazzi, A., Bomfin, R., Giakoumidis, N., Rappaport, T. S., & Chafii, M. (2025). Statistical and Deterministic RCS Characterization for ISAC Channel Modeling. arXiv E-prints, arXiv–2502.
  30. [30]Zhang, J., Lin, J., Tang, P., Zhang, Y., Xu, H., Gao, T., Miao, H., Chai, Z., Zhou, Z., Li, Y., et al. (2023). Channel measurement, modeling, and simulation for 6G: A survey and tutorial. https://arxiv.org/abs/2305.16616
  31. [31]Zhang, Y., Zhang, J., Zhang, J., Pei, Y., Liu, Y., Tian, L., Jiang, T., & Liu, G. (2025). Research and experimental validation for 3GPP ISAC channel modeling standardization. Digital Communications and Networks.
  32. [32]Gentile, C., Papazian, P. B., Sun, R., Senic, J., & Wang, J. (2018). Quasi-deterministic channel model parameters for a data center at 60 GHz. IEEE Antennas and Wireless Propagation Letters, 17(5), 808–812.
  33. [33]Varshney, N., Wang, J., Lai, C., Gentile, C., Charbonnier, R., & Corre, Y. (2021). Quasi-deterministic channel propagation model for an urban environment at 28 GHz. IEEE Antennas and Wireless Propagation Letters, 20(7), 1145–1149.
  34. [34]Gentile, C., Senic, J., Bodi, A., Berweger, S., Caromi, R., & Golmie, N. (2024). Context-aware channel sounder for AI-assisted radio-frequency channel modeling. 2024 18th European Conference on Antennas and Propagation (EuCAP), 1–5.
  35. [35]Fan, W., Wei, N., Xi, R., Bazzi, A., Xiu, Y., Assi, C., Dong, J., & Jin, J. (2026). Heterogeneous mixture-of-experts for energy-efficient multimodal ISAC in highly mobile networks. https://arxiv.org/abs/2604.06697
  36. [36]Fan, W., Wei, N., Bazzi, A., Xi, R., Song, Z., Li, Y., Zeng, Z., Xiu, Y., & Assi, C. (2026). Optimizing Tracking Accuracy in Energy-Constrained Multimodal ISAC via Lyapunov-Driven Heterogeneous Mixture-of-Experts. https://arxiv.org/abs/2604.26330
  37. [37]National Institute of Standards and Technology (NIST). (n.d.). NextG Channel Model Alliance: Data Repository. In https://nextg.nist.gov. Retrieved https://nextg.nist.govnextg.nist.gov
  38. [38]Azim, A. W., Bazzi, A., Bomfin, R., & Chafii, M. (2025). 3GPP-Compliant Radar Cross Section Characterization of Indoor Factory Targets. https://arxiv.org/abs/2505.08754
  39. [39]Manfredi, G., Russo, P., De Leo, A., & Cerri, G. (2020). Efficient simulation tool to characterize the radar cross section of a pedestrian in near field. Progress In Electromagnetics Research C, 100, 145–159.
  40. [40]Nataraja, N. K., Ali, K., Bai, F., & Molisch, A. F. (2025). Cars and Trucks as Blockers and Reflectors in Vehicular ISAC Systems: Theory, Measurements, and Applications. IEEE Vehicular Technology Magazine.
  41. [41]National Institute of Standards and Technology (NIST) Communications Technology Laboratory. (n.d.). NIST Target Channel Model for ISAC simulations. In https://github.com/usnistgov/nist-target-channel-realization. Retrieved https://github.com/usnistgov/nist-target-channel-realizationgithub.com/usnistgov/nist-target-channel-realization

Paper details

vlnc:W_evsjyjzr
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents