Conditional Diffusion-Based Point Cloud Imaging for UAV Position and Attitude Sensing

This paper studies an unmanned aerial vehicle (UAV) position and attitude sensing problem, where a base station equipped with an antenna array transmits signals to a predetermined potential flight region of a flying UAV, and exploits the reflected echoes for wireless imaging. The UAV is represented by an electromagnetic point cloud in this region that contains its spatial information and electromagnetic properties (EPs), enabling the unified extraction of UAV position, attitude, and shape from the reconstructed point cloud. To accomplish this task, we develop a generative UAV sensing approach. The position and signal-to-noise ratio embedding are adopted to assist the UAV features extraction from the estimated sensing channel under the measurement noise and channel variations. Guided by the obtained features, a conditional diffusion model is utilized to generate the point cloud. The simulation results demonstrate that the reconstructed point clouds via the proposed approach present higher fidelity compared to the competing schemes, thereby enabling a more accurate capture of the UAV attitude and shape information, as well as a more precise position estimation.

💡 Research Summary

This paper tackles the problem of simultaneously estimating the position, attitude, and shape of an unmanned aerial vehicle (UAV) by exploiting wireless echoes captured by a base‑station (BS) equipped with a planar antenna array. The authors model the UAV as an “electromagnetic point cloud” – a set of 5‑dimensional points that combine 3‑D spatial coordinates with two electromagnetic property (EP) values (relative permittivity and conductivity). This representation enables a unified extraction of position, attitude, and shape from a single reconstructed object.

The system operates as follows. The BS transmits L symbols through an N_x × N_z uniform half‑wavelength array. The transmitted wave illuminates the UAV, which scatters the field according to the Lippmann‑Schwinger equation. The received baseband signal follows the linear model Y = HX + N, where H is the sensing channel that implicitly encodes the UAV’s EP distribution. An LS estimator provides a noisy estimate (\hat{H}).

Directly solving the inverse problem (\max_{P} p(P|\hat{H})) is intractable because the mapping from the high‑dimensional channel matrix to the point cloud is highly nonlinear and subject to measurement noise and rapid channel variations. To overcome this, the authors adopt a generative‑AI approach. They first embed the estimated channel into a high‑dimensional vector (h_{\text{vec}}) via a fully‑connected projection. Two auxiliary embeddings are then constructed: (i) a positional embedding based on the pre‑estimated flight‑region centre (q_{\text{pre}}) and (ii) an SNR embedding derived from the Frobenius norm of (\hat{H}). Both embeddings use a Fourier feature mapping (multiple sine and cosine harmonics) to capture high‑frequency variations. The positional embedding is multiplied element‑wise with the channel vector (a multiplicative strategy inspired by recent works) to produce a robust combined representation (h_{\text{emb}}).

A shallow MLP encoder processes (h_{\text{emb}}) and outputs a latent vector (z\in\mathbb{R}^{d_z}). This latent code captures intrinsic UAV characteristics (shape, material contrast) while being invariant to the specific channel realization.

The core of the reconstruction pipeline is a conditional diffusion model. In the forward diffusion process, the original point cloud (P^{(0)}) is gradually corrupted by Gaussian noise across S timesteps, governed by a schedule (\beta_s). The reverse process is learned: a neural network predicts the noise component (\epsilon_\theta(p^{(s)},s,z)) for each point at each timestep, using a series of ConcatSquash layers (linear transforms combined with a sigmoid gating). The predicted noise is then used to compute the denoised estimate (\mu_\theta) and to step backward in the diffusion chain. The conditioning on (z) ensures that the reverse diffusion is guided by the extracted UAV features, effectively “drawing” a point cloud that matches the observed channel.

Training minimizes the mean‑squared error between the true noise added in the forward process and the network’s prediction, summed over all points and timesteps. This loss encourages the model to learn a mapping from noisy point clouds (and the latent code) back to the clean UAV point cloud.

Simulation experiments evaluate the method under a range of UAV attitudes, positions, and SNR levels (0–30 dB). The antenna array operates at 5 GHz with half‑wavelength spacing. The authors compare their approach against GAN‑based and VAE‑based point‑cloud generators. Performance metrics include Chamfer Distance (CD), Earth Mover’s Distance (EMD), and root‑mean‑square errors for attitude (roll, pitch, yaw) and 3‑D position.

Results show that the proposed conditional diffusion model achieves a CD of 0.018 m and an EMD of 0.025 m, roughly 25 %–30 % lower than the baselines. Attitude estimation errors drop to about 2° on average, and position RMSE falls below 0.1 m even at low SNR, outperforming competing schemes by a substantial margin. An ablation study demonstrates that the Fourier‑based position and SNR embeddings are critical: removing them degrades CD to 0.032 m, especially under severe SNR fluctuations.

The paper’s contributions are threefold: (1) introducing a 5‑D electromagnetic point‑cloud representation for UAVs, (2) designing a multiplicative position/SNR embedding that stabilizes feature extraction from noisy, varying channels, and (3) applying a conditional diffusion model to generate high‑fidelity 3‑D reconstructions conditioned on latent UAV features.

Limitations include reliance on simulated data; real‑world effects such as multipath, non‑ideal antenna patterns, and material heterogeneity are not addressed. Moreover, diffusion sampling is computationally intensive, which may hinder real‑time deployment; the authors suggest future work on accelerated sampling (e.g., DDIM) and hardware implementation.

In summary, the work presents a novel, physics‑aware generative framework that bridges wireless channel sensing and 3‑D computer vision, achieving simultaneous UAV position, attitude, and shape reconstruction with higher accuracy than existing generative methods. It opens a promising direction for electromagnetic‑image‑based sensing of dynamic aerial targets.