Physics-Inspired Target Shape Detection and Reconstruction in mmWave Communication Systems

The integration of sensing and communication (ISAC) is an essential function of future wireless systems. Due to its large available bandwidth, millimeter-wave (mmWave) ISAC systems are able to achieve high sensing accuracy. In this paper, we consider the multiple base-station (BS) collaborative sensing problem in a multi-input multi-output (MIMO) orthogonal frequency division multiplexing (OFDM) mmWave communication system. Our aim is to sense a remote target shape with the collected signals which consist of both the reflection and scattering signals. We first characterize the mmWave’s scattering and reflection effects based on the Lambertian scattering model. Then we apply the periodogram technique to obtain rough scattering point detection, and further incorporate the subspace method to achieve more precise scattering and reflection point detection. Based on these, a reconstruction algorithm based on Hough Transform and principal component analysis (PCA) is designed for a single convex polygon target scenario. To improve the accuracy and completeness of the reconstruction results, we propose a method to further fuse the scattering and reflection points. Extensive simulation results validate the effectiveness of the proposed algorithms.

💡 Research Summary

**
The paper tackles a core challenge of future 6G systems: how to simultaneously provide high‑rate communication and high‑precision environmental sensing (ISAC) using millimeter‑wave (mmWave) technology. The authors focus on a multi‑base‑station (BS) collaborative scenario in a MIMO‑OFDM system and aim to reconstruct the shape of a remote object from the received signals, which contain both reflection and scattering components.

Physical propagation model – The work begins by establishing a deterministic channel model that captures the physics of mmWave propagation on object surfaces. Using the Lambertian scattering model, the authors derive closed‑form expressions for the power of a scattered ray as a function of incidence angle, scattering angle, micro‑facet area, and attenuation coefficient. Reflection is modeled with the classic mirror‑image method. Energy conservation imposes the constraint α_r² + α_s² = 1, linking the reflection and scattering attenuation factors. This model provides per‑path angle‑of‑arrival (AoA), angle‑of‑departure (AoD), delay, and gain, which are unavailable in purely statistical models.

Signal processing pipeline – The received channel tensor H (N_r × N_t × N_c) is first transformed to the AoA/AoD/Delay domain via three‑dimensional FFT/IFFT, yielding a periodogram S = ‖F‖². The periodogram offers a fast, coarse view of the spectral peaks but suffers from limited resolution. To achieve super‑resolution, the authors apply MUSIC (Multiple Signal Classification) on two‑dimensional slices of H that correspond to each detected AoA. For each AoA, a covariance matrix is formed across the transmit antennas (R_φ^t) and across subcarriers (R_φ^c). With K = 1 (single dominant path per AoA), MUSIC extracts precise AoD and delay values, replacing the simple peak‑search in the periodogram.

Reflection points, being stronger and sparser, are identified by locating the highest periodogram peak. If a reflection exists, three separate covariance matrices (receive, transmit, subcarrier) are built and MUSIC is again applied to obtain its AoA, AoD, and delay, from which the geometric location of the mirror source is derived.

Shape reconstruction – The set of scattering points is fed into a two‑stage reconstruction algorithm. First, a Hough Transform converts each point into a sinusoidal curve in the ρ‑θ space; the accumulator is discretized into N_ρ‑θ bins. Bins with the most votes define candidate point groups (P_h^i). Because discretization limits accuracy, each candidate group is refined with Principal Component Analysis (PCA). PCA yields the major axis (line l(P)) and minor axis; two validity tests are performed: (1) uniformity of the projection of points onto the line (hypothesis test against a uniform distribution) and (2) a variance ratio σ₁/σ₂ exceeding a threshold γ_v. Only lines satisfying both are kept. Overlapping candidate groups are merged if the overlap ratio exceeds γ_l and the angular difference is below ε_θ; otherwise, a new line is added when overlap is below γ_u. The result is a set of K lines {l_s^k} that approximate the edges of a convex polygon and the associated point clusters.

Second, the previously detected reflection points are used to refine the polygon. Because a reflection point encodes the surface normal of the object, its geometric relationship with the polygon edges can be exploited to (i) add missing edges, (ii) correct edge positions, or (iii) validate the polygon orientation. This fusion step significantly improves both the completeness (all edges present) and the geometric accuracy (lower edge‑position error).

Performance evaluation – Extensive Monte‑Carlo simulations are conducted with varying numbers of BSs (single vs. dual), signal‑to‑noise ratios (0–30 dB), and polygon shapes (triangles, quadrilaterals, pentagons). Metrics include average point‑position error, polygon‑area error, and edge‑count correctness. Compared with a baseline that uses only the periodogram, the proposed periodogram + MUSIC pipeline reduces position error by roughly 40 % across all SNRs. Adding the Hough‑PCA reconstruction further cuts the area error, and the final reflection‑fusion step yields an additional ~30 % reduction, achieving sub‑decimeter accuracy in realistic bandwidth (1 GHz) and antenna array sizes (8 × 8). Computational complexity remains modest: the dominant cost is the FFT (O(N log N)) and a single‑source MUSIC eigen‑decomposition per AoA, both feasible for real‑time processing.

Conclusions and outlook – The paper demonstrates that a physics‑driven channel model combined with classical radar signal‑processing (periodogram, MUSIC) and computer‑vision techniques (Hough Transform, PCA) can enable accurate shape sensing in mmWave ISAC systems. The methodology is scalable to larger networks, and the authors suggest extensions to three‑dimensional objects, non‑convex geometries, and experimental validation with real hardware as future work.