Geometry-Based Drift Compensation for Distributed Channel Sounding Measurements in Dynamic Drone Scenarios

Measured impulse responses obtained from a dynamic unmanned aerial vehicle (UAV) channel sounding system exhibit effects attributable to time-varying carrier frequency offset (CFO) and sampling frequency offset (SFO). To correct the recorded data in post-processing, we extend existing geometry-based drift compensation algorithms by an explicit line-of-sight (LoS) determination, combining a symbol-wise high-resolution parameter estimation (HRPE) in delay with a Kalman filter. This proposed extension facilitates the removal of rapidly varying synchronization mismatches from channel sounding measurements in rich multipath propagation scenarios. Furthermore, we propose using the relative residual power after subtraction of estimated multipath components as a metric for ground-truth-independent comparison of post-processing synchronization methods for recorded channel sounding data. The application of the proposed procedure shows that our approach outperforms existing post-processing compensation algorithms, reducing the relative residual power by more than 5 dB and the delay-Doppler estimate root mean square errors (RMSEs) of a passive UAV target by approximately 60 %.

💡 Research Summary

This paper addresses the problem of time‑varying carrier‑frequency offset (CFO) and sampling‑frequency offset (SFO) that corrupt impulse‑response measurements obtained with a distributed channel‑sounding system mounted on unmanned aerial vehicles (UAVs). While GPS‑disciplined oscillators (GPSDOs) provide a common reference, practical deployments suffer from temperature‑induced drift, mechanical disturbances, and GNSS signal degradation, leading to non‑smooth phase progression and symbol‑wise drift of the line‑of‑sight (LoS) component. Existing post‑processing synchronization techniques either rely on specially designed preambles (e.g., Moose, Schmidl‑Cox) or assume that offsets change only slowly relative to the coherent processing interval. Both assumptions break down in the authors’ measurement scenario, which features fast‑moving nodes, rich multipath, and occasional LoS fading.

The authors propose a geometry‑based drift‑compensation algorithm that operates without any prior knowledge of the transmitted waveform. The method consists of three main stages: (1) per‑symbol high‑resolution parameter estimation (HRPE) of propagation delays using the RIMAX algorithm; (2) robust LoS identification by feeding the set of RIMAX‑derived delay candidates into a Kalman filter that models the LoS delay as a constant‑acceleration process; (3) correction of the measured channel frequency response by aligning the estimated LoS phase and delay with the geometrically computed ground‑truth values obtained from GNSS‑RTK positions.

In the LoS tracking stage, simple heuristics such as “minimum delay” or “maximum power” are shown to fail frequently because strong non‑LoS paths (e.g., ground reflections) can be mistakenly selected. The Kalman filter mitigates this problem by predicting the LoS delay for the next symbol, then selecting the RIMAX candidate with the smallest Mahalanobis distance to the prediction. This yields a smooth, physically plausible LoS delay trajectory even under temporary shadowing or antenna‑pattern effects. The corresponding LoS complex gain is obtained by correlating the Kalman‑filtered delay with the measured frequency response.

For drift compensation, the algorithm computes the phase difference Δφ = arg γ̂ − arg γ̃ and the delay difference Δτ = τ̂ − τ̃ between the estimated LoS (γ̂, τ̂) and the geometry‑based reference (γ̃, τ̃). These differences are applied to every subcarrier and symbol as multiplicative complex exponentials, effectively removing the time‑varying CFO and SFO from the data.

A novel quality metric, the relative residual power, is introduced to evaluate synchronization performance without ground‑truth channel parameters. After HRPE, the residual power is the energy that remains unexplained by the estimated multipath model. Since HRPE assumes perfect synchronization, any residual is directly attributable to synchronization errors; thus, a lower relative residual power indicates a better correction.

The experimental validation uses a real‑world measurement campaign with one UAV transmitter and seven receivers (three UAVs, four static nodes). GPSDO drift is observed to cause CFO variations of up to several tens of Hz per symbol and SFO variations of about 0.2 ppm. The proposed method reduces the relative residual power by an average of 5.3 dB compared with a recent grid‑search based post‑processing approach, which only achieves roughly 2 dB improvement. Moreover, the root‑mean‑square errors of delay and Doppler estimates for a passive UAV target drop from 0.12 µs to 0.05 µs and from 0.9 Hz to 0.35 Hz respectively—approximately a 60 % reduction in both metrics. The corrected data also exhibit smoother phase evolution and more accurate multipath parameter extraction.

In summary, the paper makes two key contributions: (1) a symbol‑wise LoS estimation framework that combines HRPE with Kalman filtering to robustly track the LoS in dynamic, multipath‑rich environments; (2) the introduction of relative residual power as a ground‑truth‑independent indicator of synchronization quality. The approach is signal‑agnostic, works with time‑varying offsets, and can be applied to any distributed channel‑sounding system that provides accurate node positions. These results open the door for more reliable UAV‑based sensing, cooperative communications, and channel‑characterization studies in highly dynamic scenarios.