Quality-Aware Denoising of Ultra-Short TDoA Measurements for 5G-NR UAV Localization

This paper proposes Adaptive Gain Exponential Smoother (AGES), a lightweight filtering technique that leverages 3GPP measurement quality reports to significantly reduce positioning errors for 5G-NR UAVs using ultra-short Time Difference of Arrival (TDoA) measurement sequences.

The Gist

🚁 The Big Problem: Flying Blind in a Concrete Jungle

Imagine you are flying a drone (UAV) over a busy city to deliver a package or help in a rescue mission. You need to know exactly where you are—down to the size of a coffee table (sub-meter accuracy)—and you need to know it right now (low latency).

Usually, drones use GPS. But in a city full of tall skyscrapers, GPS signals bounce off buildings (like an echo in a canyon) or get blocked entirely. It's like trying to hear a whisper in a noisy stadium; the signal gets messy, and your drone might think it's on the roof when it's actually on the street.

To fix this, we use the 5G network (the same one on your phone) to help the drone find its way. The cell towers send out special "ping" signals. The drone listens to these pings from different towers and calculates how long it took for the signal to arrive. By comparing the tiny time differences, it can triangulate its position. This is called TDoA (Time Difference of Arrival).

⏱️ The Catch: The "Speed Limit" of Accuracy

Here is the tricky part: To get a super-accurate position, the drone usually needs to listen to many pings over a long time. But in a fast-moving city, waiting too long is dangerous. If the drone waits 5 seconds to calculate its position, it might have already crashed into a building by the time the math is done.

So, the drone is forced to make a decision based on ultra-short data—maybe just 3 to 5 pings.

• The Analogy: Imagine trying to guess the speed of a car by looking at it for only a split second. If you only have one or two snapshots, it's very hard to tell if the car is speeding up, slowing down, or just driving straight. Standard math tools (like the ones used for GPS) usually need a long video clip to work well. When you give them only a split-second snapshot, they get confused and make big mistakes.

🛠️ The Solution: The "Smart Filter" (AGES)

The authors of this paper, Zexin Fang and his team, invented a new tool called AGES (Adaptive Gain Exponential Smoother). Think of AGES as a super-smart noise-canceling headphone specifically designed for this split-second situation.

Here is how it works, using a simple metaphor:

1. The "Quality Report" (The Trust Meter)

In the 5G network, every time the drone hears a ping, the network also sends a little note saying, "Hey, this signal was clear," or "Hey, this signal was a bit fuzzy."

• Old methods: Ignore the note. They treat a fuzzy signal the same as a clear one.
• AGES: Reads the note. If the signal is fuzzy, it says, "I'll trust this one a little less." If it's clear, it says, "I'll trust this one a lot."

2. The "Weighted Average" (The Recent Memory)

Since the drone is moving fast, the most recent pings are usually the most important.

• Old methods: Might try to average all the pings equally, or try to guess the drone's speed (which is hard with so little data).
• AGES: Uses a "sliding window." It gives the most recent pings the most weight and the older ones less weight. It's like remembering a conversation: you remember what the person just said much better than what they said 10 minutes ago.

3. The "Adaptive Gain" (The Volume Knob)

This is the magic sauce. AGES has a volume knob that turns up or down automatically.

• If the signals are very noisy (bad weather, tall buildings), the knob turns down, and AGES relies more on the "trust meter" (the quality report) to smooth things out.
• If the signals are clean, the knob turns up, and AGES lets the new data in quickly so the drone doesn't lag behind.

📊 What Happened in the Test?

The researchers simulated a drone flying through a city at different speeds and heights. They compared their new AGES filter against other common methods (like simple averaging or median filters).

• The Result: With only 3 to 5 pings (a very short time), AGES reduced the positioning error by 30% to 40% compared to the other methods.
• The Analogy: Imagine you are trying to aim a dart at a moving target.
  • Standard methods are like a person guessing where the target will be based on a blurry photo. They often miss.
  • AGES is like a person who not only sees the photo but also knows how shaky the camera was, and adjusts their aim instantly. They hit the bullseye much more often, even with very little information.

🏁 Why Does This Matter?

This paper proves that we don't need to wait for "perfect" conditions or huge amounts of data to fly drones safely in cities. By using a lightweight, smart filter that listens to the 5G network's own quality reports, we can:

1. Save lives: Rescue drones can navigate narrow streets without crashing.
2. Save money: We don't need expensive new hardware; this works with the 5G towers we already have.
3. Move faster: Drones can fly faster because they can calculate their position almost instantly.

In short: The authors built a "smart brain" for 5G drones that lets them find their way accurately, even when they only have a split second to look at the world.

Technical Summary

1. Problem Statement

The paper addresses the critical challenge of achieving sub-meter positioning accuracy for Uncrewed Aerial Vehicles (UAVs) in dense urban environments using 5G New Radio (5G-NR) networks, specifically under stringent latency constraints.

• The Conflict: High-accuracy localization typically requires processing long sequences of measurements to model velocity and filter noise (e.g., via Kalman filtering). However, safety-critical UAV operations (e.g., emergency delivery) require end-to-end positioning latencies of $\le 100$ ms (and ideally 10 ms).
• The Constraint: To meet these latency targets, the system must process ultra-short measurement sequences (typically only 3–7 Time Difference of Arrival (TDoA) samples) before reporting to the Localization Management Function (LMF).
• The Gap: Conventional denoising methods (like standard Kalman filters or double exponential smoothing) fail in this regime because they either lack sufficient data to estimate motion trends or introduce unacceptable delays. Furthermore, existing lightweight filters often ignore the measurement quality reports provided by the 3GPP framework.

2. Methodology

The authors propose a novel, lightweight filtering technique called the Adaptive Gain Exponential Smoother (AGES).

A. System Context

• Framework: The study utilizes the 3GPP 5G-NR Downlink Observed Time Difference of Arrival (DL-OTDOA) architecture.
• Signal Flow: User Equipment (UE/UAV) receives Positioning Reference Signals (PRS) from multiple gNBs, calculates TDoA, and reports measurements along with Quality Reports (QR) to the LMF.
• Channel Model: The simulation uses an Air-to-Ground (A2G) channel model (Urban Micro-Aerial Vehicle scenario) where path loss and Line-of-Sight (LOS) probability are functions of UAV altitude and horizontal distance.

B. The AGES Algorithm

AGES is designed to denoise TDoA measurements using only a few samples (3–7) by integrating exponentially weighted averaging with adaptive gains derived from 3GPP quality metrics.

1. Quality-Aware Variance Estimation:

   • Instead of assuming uniform noise, AGES extracts the Signal-to-Noise Ratio (SNR) and bandwidth from the 3GPP measurement reports.
   • It calculates the measurement variance ($\sigma^2_d$) using a theoretical bound derived from the Cramér-Rao Lower Bound (CRLB):
     $$\sigma^2_d \propto \frac{c}{8\pi^2 \cdot \text{SNR} \cdot \beta \cdot \ln(\beta)}$$
   • This creates a time-varying measurement covariance matrix ($R$) that reflects the actual reliability of each sample.

2. Adaptive Gain Mechanism:

   • Unlike standard Exponential Smoothing which uses a fixed decay factor, AGES computes a Kalman-style gain ($K_{gain}$) for the current measurement.
   • The gain is determined by the ratio of the accumulated historical variance to the sum of historical variance and the current measurement variance.
   • Logic: If the current measurement has high quality (low variance), the gain increases, allowing the filter to trust the new data more. If the quality is poor, the gain decreases, relying more on the weighted history.

3. Algorithm Steps:

   • Input: Distance measurements and Quality Reports.
   • Compute variance ($R$) for each sample based on SNR and bandwidth.
   • Calculate a weighted average of previous $K-1$ samples.
   • Update the current estimate using the adaptive gain and the new measurement.

3. Key Contributions

• AGES Filter: Introduction of a lightweight, quality-aware filter specifically tailored for ultra-short TDoA sequences where traditional motion modeling is impossible.
• 3GPP Integration: The method uniquely leverages existing 3GPP measurement quality reports (SNR, bandwidth) to dynamically adjust filtering weights, requiring no additional signaling overhead.
• Latency-Accuracy Trade-off: Demonstrates a solution that maintains sub-meter accuracy while adhering to the strict latency requirements of industrial IoT and UAV operations (processing only 3–5 samples).
• Comprehensive Benchmarking: The paper provides a rigorous comparison against Exponential Smoothing, Double Exponential Smoothing, Median Filtering, and Savitzky-Golay filters under various UAV altitudes and velocities.

4. Results

Simulations were conducted in a gNB-dense urban environment with UAV velocities ranging from 50 to 90 km/h and altitudes of 20–30 meters.

• Performance Gain: AGES achieved a 30–40% reduction in positioning error compared to raw measurements and other denoising techniques when using only 3–5 repeated measurements.
• Comparison with Baselines:
  • Median Filter: Effective only for very short windows where UAV displacement is negligible; performance degrades rapidly as the window expands because it ignores motion trends.
  • Double Exponential Smoothing: Underperformed with short sequences because it failed to accurately estimate trends with insufficient data points.
  • Savitzky-Golay: Showed degradation with short windows but could theoretically outperform AGES with much longer windows (at the cost of violating latency constraints).
• Robustness: AGES maintained superior performance across different altitudes (20m vs. 30m) and velocities (50 km/h vs. 90 km/h). It adapted well to varying channel conditions (LOS/NLOS probabilities) by utilizing the quality reports.

5. Significance

This work bridges a critical gap in 5G-NR localization for UAVs:

1. Enables Autonomous Operations: By enabling reliable sub-meter accuracy with minimal latency, AGES supports safety-critical urban UAV applications (e.g., medical delivery, search and rescue) where GNSS is unreliable due to urban canyons.
2. Infrastructure Compatibility: The solution is fully compatible with existing 3GPP Release 17/18 frameworks and does not require hardware upgrades or new signaling protocols.
3. Efficiency: It proves that high-precision localization does not necessarily require long observation windows if the filtering algorithm is "quality-aware," offering a practical path forward for real-time UAV navigation in complex environments.