Experimental Study of Outdoor UAV Localization and Tracking using Passive RF Sensing

Udita Bhattacherjee, Ender Ozturk, Özgür Özdemir, Ismail Güvenç, Mihail L. Sichitiu, Huaiyu DaiWiNTECH 2021 · DOI

TL;DR

This paper tackles the challenge of accurately locating and tracking Unmanned Aerial Vehicles (UAVs) using passive radio frequency (RF) sensing, a crucial task for airspace security. The authors propose an Extended Kalman Filter (EKF) framework that segments the UAV's trajectory and applies different motion models to each segment, significantly improving localization accuracy compared to raw RF sensor data.

What Problem Does This Paper Address?

The increasing prevalence of UAVs, while offering numerous benefits, also raises significant privacy and security concerns. Effectively detecting, locating, and tracking these aerial vehicles is paramount for maintaining safe and secure airspace. Traditional methods like radar can be challenging for small UAVs due to their low radar cross-section. This paper focuses on passive RF sensing, which leverages the RF signals emitted by UAVs. However, existing RF sensing localization techniques, even with commercial systems like Keysight's N6854A, can suffer from significant localization errors. This paper aims to bridge this gap by proposing an improved localization and tracking method using an EKF that accounts for the dynamic and often unpredictable motion of UAVs.

Methods

The core of this paper's contribution lies in its Motion Model Based Extended Kalman Filter (EKF). The authors acknowledge that UAV trajectories are often complex and cannot be accurately represented by a single motion model (like constant velocity or constant acceleration) throughout their flight. To address this, they propose a multi-motion model approach:

  • Trajectory Segmentation: The UAV's flight path is divided into multiple segments.
  • Motion Model Assignment: For each segment, an appropriate motion model is chosen. The paper considers three common kinematic models:
  • Constant Velocity (CV-MM): Assumes constant speed and direction (straight line).
  • Constant Acceleration (CA-MM): Assumes constant acceleration.
  • Constant Turn (CT-MM): Assumes movement along a circular path with constant angular velocity.
  • EKF Application: An EKF is applied independently to each segment, utilizing the best-fitting motion model for that specific part of the trajectory. This allows the filter to adapt to the changing dynamics of the UAV's movement.

The authors also discuss the importance of selecting appropriate noise covariance parameters (process noise $Q_k$ and measurement noise $R_k$) for the EKF, treating them as design parameters influenced by the chosen motion model.

Datasets

The study utilizes a realistic experimental dataset collected during an outdoor flight experiment.

  • UAV: A DJI Inspire 2 drone was used, operating in the 2.400-2.483 GHz band.
  • RF Sensors: Four Keysight N6841A RF sensors were deployed to detect and record the UAV's RF signals.
  • Ground Truth: The UAV's own GPS data, recorded at 100 msec intervals, served as the ground truth for evaluating localization accuracy.
  • Data Processing: The collected RF sensor data was post-processed using Keysight's N6854A geolocation software, which employed a Time Difference of Arrival (TDOA) based localization algorithm. The data was then synchronized and filtered to create a consistent dataset for analysis.

The paper does not introduce a new benchmark dataset but rather uses a real-world experimental setup to validate their proposed method.

Key Insights and Ideas

  • Dynamic Trajectories Require Dynamic Models: The most significant insight is that accurately tracking a maneuvering UAV necessitates adapting the localization model to its changing motion. A single, static motion model is insufficient for complex, real-world trajectories.
  • Segmentation is Key: Dividing the trajectory into segments and applying specific motion models to each segment is crucial for improving EKF performance. This allows the filter to better capture the nuances of the UAV's flight.
  • EKF Improves High-Error Portions: The EKF framework, particularly when combined with appropriate motion models, is effective at reducing the highest localization errors, which are often the most problematic for practical applications.
  • Model-MM Alignment Matters: The performance of the EKF is highly dependent on how well the chosen motion model aligns with the actual UAV movement within a segment. When there's a good match, error statistics significantly improve. Conversely, misalignment can lead to degraded performance.
  • Parameter Tuning is Critical: The choice of noise covariance parameters ($Q_k$ and $R_k$) significantly impacts EKF performance. While the paper uses fixed values based on experimental data, it highlights the potential for further optimization through adaptive parameter selection.

Key Results

The study demonstrates a significant improvement in localization accuracy when using the proposed EKF framework compared to raw RF sensor data.

  • Error Reduction: For "clean data" (errors below 60m), the EKF consistently reduced the maximum, mean, and standard deviation of localization errors across various motion segments.
  • Specific Segment Improvements: For instance, in segment S1 (CT-MM), the mean error was reduced from 8.0m (RF) to 5.89m (EKF), and the standard deviation from 3.49m to 2.28m. Segment S7 (CA-MM) showed particularly strong performance, with all error statistics being reduced by the EKF.
  • Raw Data Handling: For "raw data," the EKF was also effective in reducing the highest errors, though it sometimes came at the cost of slightly increased errors in other parts of the data due to the high variability in raw measurements.
  • Overall: The EKF approach successfully mitigated the highest localization errors, making the system more robust for practical applications.

Limitations and Future Work

  • Segmentation Accuracy: The paper acknowledges that perfect, real-time trajectory segmentation for unplanned trajectories is a challenging problem. In this study, segmentation was performed manually based on visual inspection and GPS data.
  • Model-MM Alignment: The performance of the EKF is sensitive to the alignment between the chosen motion model and the actual UAV movement. Misalignment can lead to performance degradation.
  • Parameter Selection: While the paper uses fixed noise covariance parameters, it suggests that these parameters could be dynamically adjusted.
  • Future Work:
    • Automated Segmentation: Developing algorithms for automatic and real-time trajectory segmentation.
  • Adaptive Parameter Tuning: Implementing machine learning techniques (e.g., reinforcement learning or ANNs) to dynamically adjust EKF parameters (noise covariances) based on the UAV's motion and environmental conditions.
  • Glitch Detection: Developing methods to detect and eliminate erroneous RF sensor data points that might originate from nearby interfering RF sources.

Summary updated

BibTeX
@inproceedings{bhattacherjee2021experimental,
  author = {Udita Bhattacherjee and Ender Ozturk and Özgür Özdemir and Ismail Güvenç and Mihail L. Sichitiu and Huaiyu Dai},
  title = {{Experimental Study of Outdoor UAV Localization and Tracking using Passive RF Sensing}},
  booktitle = {WiNTECH},
  year = {2021},
  doi = {10.1145/3477086.3480832},
}
Figure