Point Cloud-Driven CKM

• Point cloud-driven CKM is a technique that leverages detailed 3D point cloud data to generate spatial channel state information maps with improved predictive accuracy.
• It combines co-focal ellipsoid region selection and neural network architectures to capture multipath phenomena and environmental material cues effectively.
• Empirical results show significantly lower RMSE in channel gain estimation compared to traditional methods, highlighting its robustness in urban microcell scenarios.

Point cloud-driven channel knowledge mapping (CKM) leverages rich 3D environmental representations to significantly improve the construction of spatial channel state information (CSI) maps for wireless systems. By integrating high-fidelity point cloud data and advanced region selection or geometry encoding algorithms, such approaches surpass traditional methods reliant on oversimplified environment models, leading to substantially enhanced predictive accuracy for key channel metrics such as the power delay profile (PDP) and received power distributions ("radio maps") (Wang et al., 26 Jun 2025, Yuan et al., 2024).

1. Overview and Rationale

CKM refers to the task of learning spatial maps that encode partial or complete channel responses—most notably, the PDP or received signal power—at arbitrary locations in an environment. These maps facilitate environment-aware communications by reducing the need for dense, costly pilot-based channel measurements. Traditional CKM construction frequently depends on coarse geometric or parametric environmental representations (e.g., 2D/3D floorplans, basic ray-tracing), which leads to notable model-experiment mismatches due to inadequate characterization of multipath phenomena, blockage, and material properties (Wang et al., 26 Jun 2025).

Point cloud-driven CKM employs detailed 3D point clouds, often acquired via UAV photogrammetry or LiDAR, to capture both the geometry and material cues (e.g., color, surface normal estimates, angle features) pertinent to radio propagation. The framework is typically joint model- and data-driven, in which physically-motivated environmental selection is combined with machine learning, particularly neural network architectures, to yield efficient and robust CKM construction.

2. Point Cloud Environmental Representation and Acquisition

High-resolution point cloud generation is the foundation of point cloud-driven CKM. Data acquisition employs UAV-mounted imaging (e.g., DJI M300 RTK with Zenmuse P1), synthesizing multi-view images into 3D meshes and finally dense point clouds through pipelines such as DJI Terra and CloudCompare (Wang et al., 26 Jun 2025). Key properties:

• Spatial coverage: e.g., 58.9 million points for urban microcell sites (~30 m × 30 m and ~30 m × 100 m) with ~2.5 cm RMSE reconstruction error.
• Per-point features: Position ($p_i\in\mathbb{R}^3$), RGB color (for material inference), surface normals, and incident/outgoing angles relative to measured transceivers.
• Preprocessing: Spatial normalization (coordinate system centering and rotation), stochastic downsampling or upsampling to fixed subset cardinality.

This granular environmental data enables causal linking between local geometry/material features and channel responses, which is essential for accurate CKM construction under real-world multipath, scattering, and shadowing conditions.

3. Spatial Region Selection via Co-Focal Ellipsoids

A principal methodological innovation is the use of co-focal ellipsoid region selection to identify subsets of the point cloud causally related to particular multipath channel components (distinguished by time-of-arrival, ToA) (Wang et al., 26 Jun 2025).

Region Formulation

• For a transmitter ($x_T$) and a candidate receiver ($x$), and given ToA interval $[t_k, t_{k+1})$ (with $d_k = c\,t_k$ as path length), region $R_k$ is the locus of points $p$ in 3D such that

$\|p-x_T\|_2 + \|p-x\|_2 = d_k$

with additional constraints for higher-order bins.

• The first-arrival region: Interior of ellipsoid ($R_0$).
• Subsequent bins: Spherical shell regions between ellipsoids ($R_k$ for $x_T$0).

Point Selector Algorithm

Algorithm I systematically applies these geometric predicates to the entire point cloud. For each ToA bin:

• Compute and apply a translation and Rodrigues' rotation, aligning $x_T$1 along the global $x_T$2-axis for efficient evaluation.
• For each point, check membership in $x_T$3.
• Output $x_T$4, the subset of points ("interacting objects", IO) causally implicated for the corresponding ToA bin.

This yields $x_T$5 region-conditioned subsets per location pair, each encoding the environmental context most likely to influence a specific portion of the multipath response.

4. Neural Channel Gain Estimation and Learning Architecture

Given $x_T$6 subsets, the next stage estimates the physical channel gain ($x_T$7) for each multipath bin using a neural channel gain estimator (Wang et al., 26 Jun 2025). The input is a $x_T$8 matrix (geometry + features), processed as follows:

• Two-level set abstraction (PointNet++ style):
  • Level 1: Farthest point sampling, local neighborhood aggregation, MLP + max-pooling for regional feature extraction.
  • Level 2: Repeat on lower-resolution abstraction.
• Global encoding: Further PointNet operations yield a global feature vector for the bin.
• Regression head: Fully-connected layers with ReLU, batch normalization, and Dropout predict channel gain $x_T$9.

Learning objective: Minimize mean squared error (MSE) between predicted and empirical channel responses, with masking for bins lacking non-empty $x$0 and floor constraints to handle measurement sensitivity.

Training employs real-world point cloud/CSI pairs (e.g., 9,300 samples, 200 epochs, stochastic gradient descent) with no additional data augmentation beyond stochastic downsampling.

5. Performance Metrics, Benchmarks, and Empirical Advantages

Point cloud-driven CKM is quantitatively evaluated against ray-tracing, Kriging interpolation, and neural radio map baselines using RMSE for both PDP and received power (dB). On the CUHK–Shenzhen urban microcell datasets (Wang et al., 26 Jun 2025):

Method	PDP RMSE (dB) Site 1	PDP RMSE (dB) Site 2	Radio Map RMSE (dB) Site 1	Radio Map RMSE (dB) Site 2
Proposed	2.95	3.84	1.04	0.59
Wireless Insite RT	7.32	8.11	2.88	1.92
Kriging Interpolation	–	–	1.68	0.99
RadioUNet	–	–	~3.5*	~2.8*

*RadioUNet after pixel-value re-scaling, with degraded performance in tree-covered regions.

Statistical testing confirms that the proposed method's RMSE is significantly lower (p < 0.01) than all baselines, reflecting notable enhancements in spatial prediction accuracy. Key contributing factors include:

• Geometry/material precision from the point cloud
• Causal selection via co-focal ellipsoids (Algorithm I)
• Neural model's ability to encode empirical EM phenomena beyond analytic modeling limitations

Computationally, the region selection is $x$1 (with spatial indexing for acceleration), and inference is less than $x$2 ms per location using GPU resources.

6. Point Cloud-Driven Continuous Kernel Mixture (CKM) Geometry Encoding

The continuous kernel mixture (CKM) formalism, with specific instantiations such as VecKM, offers a generic, efficiency-oriented scheme for encoding local point cloud geometry for downstream tasks (Yuan et al., 2024). For a central point $x$3,

$x$4

VecKM vectorizes this mixture using random Fourier features, yielding descriptors that are provably equivalent to the Gaussian kernel mixture, preserve geometric similarity, and support reconstructability of point sets from their encodings as $x$5.

Dense encoding leverages a factorized computation, eliminating explicit neighborhood grouping, and enabling fully linear $x$6 scaling in both time and space for $x$7 points, code dimension $x$8, and auxiliary Fourier parameter $x$9. This property supports large-scale 3D environments encountered in wireless channel modeling.

Empirically, VecKM achieves superior accuracy and efficiency to PointNet, PointNet++, and point transformer baselines in normal estimation, classification, and segmentation tasks, supporting downstream applications in CKM—a plausible implication is robust non-local aggregation for fine-grained radio environment modeling. However, hyperparameter selection (kernel radii, feature dimension) requires careful tuning.

7. Future Prospects and Extensions

Future directions for point cloud-driven CKM include:

• Dynamic environments: Integrating temporal updates or moving-object detection in the point cloud to address channels with time-varying scatterers (Wang et al., 26 Jun 2025).
• Multi-frequency fusion: Leveraging both LiDAR (for millimeter-wave) and photogrammetric data to handle frequency-dependent propagation effects.
• Physics-guided learning: Incorporation of ray-tracing outputs or analytical channel model priors as neural network inputs or loss regularizers for hybrid model/data-driven learning.
• Advanced CKM geometry encoders: Adapting VecKM for anisotropic or learned kernels, multi-scale mixtures, or spatio-temporal encoding to capture nonstationary propagation conditions (Yuan et al., 2024).

These advances aim to further bridge the gap between environmental sensing fidelity and real-world wireless system requirements, cementing point cloud-driven CKM as a foundational technique for next-generation environment-aware communications.

---

References:

• "Point Cloud Environment-Based Channel Knowledge Map Construction" (Wang et al., 26 Jun 2025)
• "A Linear Time and Space Local Point Cloud Geometry Encoder via Vectorized Kernel Mixture (VecKM)" (Yuan et al., 2024)