PaperlayerView exact revision on arXiv

Source-linked AI summary

ISAC Imaging by Channel State Information using Ray Tracing for Next Generation 6G

Ahmad Bazzi, Mingjun Ying, Ojas Kanhere, Theodore S. Rappaport, Marwa Chafii

arXiv:2509.06672v1eess.SP

TL;DR

CSI-based ISAC imaging must represent complex multipath, including paths that do not follow single-bounce geometry, to reconstruct wireless environments. This paper uses NYURay-generated per-path CSI at 6.75 GHz, optimizes two TX/RX path segments into equivalent reflection points, and fuses multiple viewpoints. The resulting reconstructions capture surfaces, edges, curved features, and multi-bounce object structure in simulated scenes.

  • Problem

    Single-bounce CSI imaging is limited for representing complex objects, while 6G applications require accurate three-dimensional environmental mapping.

  • Method

    The framework converts CSI multipath angle-delay information into equivalent reflection points through two-segment optimization and fuses ERPs from multiple TX-RX positions.

  • Results

    The framework reconstructs object surfaces, edges, curved features, and dense three-dimensional point clouds across multiple simulated object scenarios.

  • Takeaways & Limitations

    Multi-vantage ERP fusion can fill single-view gaps and produce near-continuous reconstructions of foliage, trunks, cubes, plates, and other object structures.

Abstract

from arXiv · show

Integrated sensing and communications (ISAC) is emerging as a cornerstone technology for sixth generation (6G) wireless systems, unifying connectivity and environmental mapping through shared hardware, spectrum, and waveforms. The following paper presents an ISAC imaging framework utilizing channel state information (CSI) per-path components, transmitter (TX) positions, and receiver (RX) positions obtained from the calibrated NYURay ray tracer at 6.75 GHz in the upper mid-band. Our work shows how each resolvable multipath component can be extracted from CSI estimation and cast into an equivalent three-dimensional reflection point by fusing its angle and delay information, which is useful and challenging for multi-bounce reflections. The primary contribution of the paper is the two-segment reflection point optimization algorithm, which independently estimates the path lengths from the TX position and RX position to an equivalent reflection point (ERP) on the object surface, thus enabling precise geometric reconstruction. Subsequently, we aggregate the ERPs derived from multiple pairs of TX and RX positions, generating dense three dimensional point clouds representing the objects in the channel. Experimental results validate that the proposed ISAC imaging framework accurately reconstructs object surfaces, edges, and curved features. To the best of our knowledge, this paper provides the first demonstration of multi bounce ISAC imaging using wireless ray tracing at 6.75 GHz.

I. INTRODUCTION

The paper develops CSI-based RF imaging for 6G ISAC, addressing the limited fidelity of single-bounce approaches by reconstructing multi-bounce propagation as equivalent reflection points. Using NYURay-generated channel information, it proposes two-segment optimization and multi-vantage fusion for 3D scene representation.

  • Motivation: CSI-based imaging can map environments because multipath phases encode fine-scale geometry, while radio waves operate through foliage, walls, and varying lighting.Prior CSI imaging located human targets and gestures but typically assumed single-bounce reflections.
  • Objective and scope: The framework converts every resolvable multipath component into an equivalent three-dimensional reflection point using angle-delay pairs.This representation is designed to handle especially challenging multi-bounce components.
  • Objective and scope: A two-segment optimization independently estimates transmitter-to-point and point-to-receiver path lengths, producing a geometrically consistent scene representation.The method uses transmitter and receiver positions together with multipath information.
  • Channel simulation: NYURay supplies site-specific, physics-based multipath parameters including delays, angles, powers, penetration, diffuse scattering, and multi-bounce paths.Its calibrated operation includes 6.75 GHz measurements and supports large bandwidths.
  • Channel simulation: Synthetic CSI generated from NYURay ground-truth rays enables algorithm testing across many TX/RX viewpoints before operation with live CSI.The generated CSI includes known per-path geometry and hardware-related effects can be distinguished from the underlying multipath structure.
  • Contributions: The paper uses NYURay paths at 6.75 GHz to provide data-driven imaging with genuine material interactions rather than idealized channel models.The contribution list highlights multiple transmitter, receiver, and object types.

A. Mathematical Description

The mathematical model represents each TX-RX channel as multiple path components described by angular, delay, and gain parameters. Because multi-bounce propagation breaks single-bounce geometric relations, the framework absorbs entire interaction chains into equivalent reflection points.

  • Channel model: At 6.75 GHz, each TX-RX channel is modeled as a superposition of K_i ≥ 1 path components, with the LoS path indexed by j = 0 when present.Paths with j > 0 interact with the object and may be weaker than the LoS component.
  • Channel model: Each path is characterized by angle of departure, zenith departure angle, angle of arrival, zenith arrival angle, time of arrival, and path gain.These six parameters form the sensing representation used by the imaging framework.
  • Array model: The transmit and receive array responses use three-dimensional steering vectors, exemplified through Kronecker products of x- and y-axis uniform linear array responses.The formulation explicitly represents both transmit and receive array steering.
  • Multi-bounce assumption: Under multi-bounce propagation, transmit angles and propagation delay no longer obey the usual single-bounce geometric relation.The resulting mismatch reflects multiple interactions in the environment.
  • Multi-bounce assumption: A multi-bounce trajectory may combine specular reflection, diffuse scattering, and edge diffraction, so the entire sequence is abstracted as one equivalent reflection point.The ERP preserves the observable equivalent delay and direction change without recovering intermediate bounce locations.

B. Ray-tracer and CSI

The ray-tracer and CSI stage connects standardized CSI representations with NYURay-generated per-path parameters. The resulting synthetic CSI supports repeated ERP computation across TX-RX drops, while current validation remains simulation-based.

  • CSI representation: CSI is the complex MIMO channel-frequency response obtained from known pilot sequences, with each subcarrier containing the channel for every TX-RX antenna pair.The definition applies to LTE and 5G NR CSI measurements.
  • CSI representation: Beamforming codebooks provide coarse AoD/AoA estimates, which can be refined with high-resolution multipath-parameter extraction methods.Examples include matrix-pencil delay and AoA estimation and deep-learning AoA/AoD estimation.
  • NYURay pipeline: NYURay generates each path’s six-tuple from a 3D model using physics-based interactions with the object.The model is rendered in Blender, and different TX-RX drops represent different TX-RX positions.
  • NYURay pipeline: For every six-tuple path component, the pipeline computes an ERP, forms an RF image for the current view, and repeats the process for subsequent TX-RX drops.The proposed technique can also use other ray tracers that provide per-path delays and directional information.
  • Simulation scope: The ray tracer limits computational cost by bounding allowed bounces and stochastically pruning very low-power paths.The paper uses these simulations to validate the imaging algorithm and reserves real hardware experiments for future work.

III. IMAGING VIA EQUIVALENT REFLECTION POINT COMPUTATIONS

The imaging computation uses path delay and directional information to constrain a scattering point relative to the known transmitter and receiver positions. The point is represented by separate distances along the outgoing and incoming path directions.

  • Equivalent reflection point computation: For each path, total propagation delay provides an approximate path length, while departure and arrival angles define transmitter-side and receiver-side unit directions.The scattering point must lie on both the transmitter-originating and receiver-approaching directional lines.
  • Equivalent reflection point computation: The candidate scattering point is parameterized as p_Tx_i + α_(i,j)d̂_T_i,j and equivalently as p_Rx_i − β_(i,j)d̂_R_i,j, with positive segment lengths.This two-sided geometric constraint supplies the basis for ERP optimization.

A. Why Directions/Delays are not Geometrically Consistent ?

Ideal single-bounce geometry imposes reflection-angle, total-path-length, and surface-location constraints, but diffuse scattering can violate the single-point model.

  • Ideal specular reflection requires equal incidence and reflection angles at the scattering point.
  • The total path length must equal the sum of transmitter-to-point and point-to-receiver distances.
  • The scattering point must lie on the reflecting surface, a condition complicated by diffuse scattering over irregular regions.

B. Formulating an Equivalent Reflection Point Problem

The paper formulates ERP estimation as a constrained least-squares problem that fits transmitter- and receiver-side path segments to measured directions and delay. The resulting effective point represents the best geometric fit for single- and multi-bounce paths.

  • ERP estimation seeks a point whose transmitter and receiver directions align approximately with measured AoD/ZoD and AoA/ZoA.
  • The least-squares objective minimizes the squared distance between transmitter- and receiver-defined parametric lines.
  • The measured delay constrains the two segment lengths through α(i,j) + β(i,j) = Li,j.
  • The optimization separately solves each path for each TX-RX pair, with bounded segment variables and KKT conditions enforcing constraints.
  • For a single-bounce path, the cost is zero and the two segments intersect uniquely; multi-bounce paths generally produce a positive residual.
  • The ERP is therefore a least-squares best-fit location for the dominant interaction of a potentially complex physical trajectory.

C. Multi-vantage fusion

Multi-vantage fusion aggregates equivalent reflection points from TX-RX pairs after filtering implausible candidates. The filtering threshold rejects paths whose inferred internal geometry is excessively long, while adaptive uncertainty-based criteria are also possible.

  • The fusion step aggregates ERPs computed across TX-RX pairs to form the reconstructed point cloud.
  • Candidates are rejected when their geometric chord length satisfies ∥a∥ ≥ γ.
  • The threshold γ represents the straight-line chord between entry and exit points of a multi-bounce path inside the object.
  • The filter targets candidates whose incoming and outgoing rays imply an unreasonably large internal path, often indicating poor optimization convergence.
  • The fixed Euclidean threshold can be replaced by an adaptive radius proportional to an uncertainty estimate from residuals or the Cramér-Rao bound.

D. Computational Complexity

ERP computation has constant cost per path because intermediate bounces are abstracted as an approximate ERP. Across TX-RX pairs, total processing scales linearly with the number of pairs and path components.

  • One ERP computation has complexity O(1) because intermediate bounces are represented by an approximate ERP.
  • For N TX-RX pairs, computing all ERPs has complexity O(N maxi Ki), which is linear in N.

IV. SIMULATION RESULTS

Simulations at 6.75 GHz evaluate the imaging framework on six Blender-rendered objects across multiple TX-RX views. Fusing views reconstructs surfaces and features, while reconstruction quality improves with additional TX-RX pairs but depends on object geometry and noise-floor constraints.

  • Test Objects: The evaluation uses six Blender-rendered 3D objects, including trees, metal cubes, planar plates, a curved plate, and a Tesla vehicle.The objects probe branched, scaled, planar, curved, and realistic vehicular geometries.
  • Standard Tree: Seven TX-RX placements fuse complementary tree views into an almost continuous mid-canopy shell and comprehensive 3D map.Individual views recover different branch, foliage, and trunk regions, reducing shadow zones through complementary perspectives.
  • Metal Cubes: Seven fused views reconstruct near-closed metallic cube surfaces at both 1 m and 4 m scales, although larger objects require wider angular coverage.The 4 m cube forms a quasi-continuous approximately 4 m-per-side reconstruction.
  • Metal Plates: For curved and triangular plates, dominant returns localize along rims or edges while weaker returns reveal planar, underside, and secondary scattering.The triangle reconstruction concentrates intense returns along two linear edges and also shows weaker planar and apex scatterers.
  • Tesla Vehicle: Fusing seven Tesla views produces a richly detailed continuous vehicle body, combining strong hood, windshield, roofline, side-panel, and wheel-arch returns.The result demonstrates reconstruction of automotive shape from limited-view measurements.
  • Noise Floor: A −100 dBm noise floor preserves some object outlines, but the car, cubes, and triangle require additional TX-RX views for reconstruction.The foliage canopy and circular arc retain their spatial outlines under the stated suppression threshold.
  • TX-RX Tradeoffs: Chamfer distance decreases as TX-RX pairs increase, reaching about 2^-11 for flat metallic surfaces after roughly six pairs, while complex objects need about 8 to 10 pairs.The standard tree shows about 28.6 linear-scale reduction, and additional pairs yield about 20.3 linear-scale improvement for flat plates.

V. OPEN CHALLENGES

The paper identifies practical barriers to deploying CSI-based ISAC imaging, including imperfect hardware and CSI, resource demands, outliers, and computational cost. It also outlines future directions involving digital-twin integration, synthetic aperture concepts, and aggregate path weighting.

  • Real-time computational efficiency: Ray tracing and reflection-point optimization demand substantial computational resources, creating a deployment challenge for real-time ISAC imaging.
  • Synthetic Aperture Radar inspired RF images: CSI collected across time and distance could synthesize a broader effective bandwidth and potentially improve RF image resolution in FR3 systems.
  • Bandwidth and SNR: The assumed perfect CSI is limited in practice by bandwidth and SNR, while bandwidth, antenna count, and movement determine path separability and image clarity.
  • Aggregate cost for imaging: An aggregate imaging cost could weight weaker paths less according to path reliability, extending the path-by-path ERP reconstruction.
  • Model mismatch and hardware impairments: Real-world phased-array phase noise, mutual coupling, and calibration errors can distort measured CSI and require joint impairment or geometry estimation.
  • Digital-twin integration: Integrating the imaging algorithm with digital twins, multimodal frameworks, and learned scene priors may improve RF images in complex environments.
  • Outlier detection of environmental points: The framework may produce environmental outliers when paths penetrate non-total-reflector objects, motivating outlier detection against environmental point clouds.

VI. CONCLUSION & FUTURE WORK

The paper presents an end-to-end 6.75 GHz ISAC imaging pipeline that converts NYURay-generated CSI per-path components into equivalent reflection points and fuses multiple views into dense 3D point clouds. Results span varied object sizes and shapes, while the authors also show recovery of standard-compliant CSI parameters.

  • The pipeline uses NYURay CSI and a physics-guided inverse solver to localize equivalent reflection points for objects of varied shapes and sizes.
  • Joint transmit- and receive-segment optimization uses path delays and angles to represent multi-bounce paths with single-point equivalent reflection points.
  • Multi-vantage fusion combines views from different TX-RX positions into dense 3D point clouds for cubes, vehicles, plates, and trees.
  • The method scales across object sizes and reconstructs planar faces, edges, and corners without assuming known surfaces.
  • The authors report that standard-compliant CSI-RS/SRS measurements can recover the same six-tuple parameters generated with NYURay.
Loading 2509.06672v1…