💻 computer science

5G ISAC-Based UAV Detection and 3-D Tracking Using Uplink Sounding Reference Signals on an End-to-End O-RAN Simulation Testbed

This paper presents an end-to-end O-RAN simulation testbed that repurposes 5G uplink Sounding Reference Signals for passive UAV detection and 3-D tracking, demonstrating that altitude observability can be achieved through either a planar receive array or a second transmitter while maintaining concurrent communication services.

Original authors: Arun K. Gurung, Satha K. Sathananthan, Shiva R. Pokhrel

Published 2026-08-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Arun K. Gurung, Satha K. Sathananthan, Shiva R. Pokhrel

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the sky above our cities is getting crowded with tiny, buzzing drones. Some are delivering pizza, others are filming movies, but a few might be spying or causing trouble. To keep us safe, we need to spot these invisible intruders, but traditional radar is like a giant, expensive spotlight that needs its own special power line and a huge budget to run. It's hard to put one on every street corner.

Enter a clever idea called "Integrated Sensing and Communication" (ISAC). Think of it like a lighthouse that does double duty: it sends out a beam to guide ships (communication) but also listens to the echoes bouncing off the water to see if a storm is coming (sensing). In this story, the "lighthouse" is our existing 5G cell network, and the "beam" is the signal our phones send up to the tower. The big question scientists are asking is: Can we turn these everyday cell signals into a giant, invisible radar net that can catch drones without building a single new tower? This paper dives right into that question, using a super-advanced computer simulation to see if it's actually possible to turn a 5G network into a drone-hunting radar.

The researchers built a complete, virtual 5G network inside a computer to test this idea. They didn't just guess; they created a digital "test track" where a fake drone flies around, and they tried to catch it using only the signals that a normal phone sends to the tower. They found that yes, it is possible to detect and track these drones in 3D space using just the existing network, but there are some tricky geometry puzzles to solve first.

Here is how they did it and what they discovered:

The Magic Trick: Listening to the Echo
Usually, a cell tower just listens to your phone to hear your voice or data. But in this experiment, the tower listened differently. They used a special signal called the "Sounding Reference Signal" (UL-SRS), which phones send up just to help the tower tune in. The researchers treated this signal like a radar ping. When the signal hit a drone, it bounced off and came back to the tower. By measuring how long the echo took to return and how the signal changed, the tower could figure out where the drone was. The best part? They didn't need to change the phone or the tower's software standards; they just repurposed a signal that was already there.

The "One-Eye" Problem and the Fix
There was a catch. With just one phone (the transmitter) and one tower (the receiver), the system could tell how far away the drone was and which direction it was going left or right, but it couldn't tell how high up it was. It was like trying to guess the height of a bird in the sky while looking at it with only one eye closed; you know it's there, but you can't tell if it's flying at 10 feet or 100 feet.

To fix this, the team tried two clever tricks, like giving the system a second pair of eyes:

The Vertical Array: They gave the tower a special antenna that was tall and flat (like a picture frame) instead of just a line. This allowed the tower to "see" up and down, not just side-to-side.
The Second Phone: They added a second phone on the other side of the street to send signals. Now, the tower had two different angles to look at the drone from. By comparing the two different paths the signals took, the math could figure out the height perfectly, just like how your two eyes help you judge depth.

The Results: Catching the Drone
The simulation showed that this system works surprisingly well.

Detection: The system could spot the drone in about 92.7% of the time it was visible.
Accuracy: When they used the second phone trick (the two-transmitter setup), the system guessed the drone's height with an error of only 2.2 meters (about 7 feet). That's pretty close for a simulation!
Speed: Even while the phone was sending data at 10 Mbit s⁻¹ (a normal internet speed), the radar system still worked perfectly. It didn't get confused by the phone's regular traffic.

The "What Ifs" and Limits
The researchers were very careful to point out what didn't work perfectly.

The Height Guess: If they tried to track the drone with just one phone and guessed the wrong height (like assuming the drone was low when it was actually high), the tracking went completely wrong. The system needs that extra "eye" (either the tall antenna or the second phone) to be accurate.
The "Street Canyon" Problem: In a simulation where tall buildings blocked the view, the system sometimes lost the drone. This makes sense; if a building blocks the signal, the radar can't see the echo.
Simulation vs. Reality: It is important to remember that all of this happened inside a computer simulation using a "ray-traced" model (a fancy way of mathematically predicting how radio waves bounce off buildings). The paper says this is a proof-of-concept. They haven't built a real-life version with real hardware yet, though they plan to.

Why This Matters
This paper proves that we might not need to build expensive, dedicated radar systems to catch drones. Instead, we could potentially use the cell towers and phones already all over the city. It's a bit like realizing your streetlights could also act as security cameras if you just looked at the reflections differently. While there are still hurdles to clear before this is a real-world product, this study shows that the idea is mathematically sound and ready for the next step: building it for real.

Technical Summary: 5G ISAC-Based UAV Detection and 3-D Tracking Using Uplink Sounding Reference Signals on an End-to-End O-RAN Simulation Testbed

Problem Statement
The proliferation of low-altitude unmanned aerial vehicles (UAVs) necessitates scalable, cost-effective counter-UAS (C-UAS) solutions. Traditional radar systems require dedicated spectrum and infrastructure, which are often prohibitive. While 5G Integrated Sensing and Communication (ISAC) offers a vision of using existing cellular infrastructure for dual-purpose sensing, practical implementations remain scarce. Specifically, there is a lack of end-to-end open testbeds that integrate a real 5G New Radio (NR) protocol stack, a standards-compliant O-RAN Near-RT RIC, and a ray-traced channel simulator to validate UAV surveillance. Furthermore, standard 3GPP sensing specifications (e.g., for Release 19) are still under study, and the uplink sounding reference signal (UL-SRS) is a candidate, not a standardized, sensing waveform.

Methodology
The authors present an end-to-end O-RAN simulation testbed built from open-source components: OpenAirInterface (OAI) for the 5G NR stack, FlexRIC for the Near-RT RIC, and Sionna RT for ray-traced channel emulation. The system repurposes the standard UL-SRS as a passive radar waveform without modifying the NR standard or requiring dedicated sensing signals.

Architecture: The testbed operates as a containerized stack. The OAI gNB hosts a custom ISAC sensing stage within its Physical Layer (PHY). The OAI nrUE acts as the illuminator, transmitting UL-SRS. The UAV is a non-cooperative passive scatterer.
Sensing Pipeline: Inside the gNB PHY, the system performs:
1. Clutter Subspace Deflation: A rank- $K_c$ subspace deflation removes the dominant static direct path and ground bounce, isolating the UAV echo.
2. Range-Doppler Processing: A 2-D OFDM Range-Doppler map is generated via zero-padded IDFT (range) and DFT (Doppler) over a Coherent Processing Interval (CPI) of 640 ms.
3. Detection: Order-Statistic CFAR (OS-CFAR) with single-peak grouping detects targets.
4. Angle Estimation: A closed-form interferometric estimator calculates the Angle of Arrival (AoA) using a 4-element Uniform Linear Array (ULA).
Transport: Detections are transmitted to a subscribing xApp in the Near-RT RIC via a custom E2 service model (SM-SENS) using standard E2AP procedures.
Tracking: The xApp runs an Extended Kalman Filter (EKF) with a nonlinear measurement model (bistatic range, range-rate, and azimuth).
3-D Tracking Solutions: Since a single bistatic pair with a 1-D array cannot observe elevation (requiring a height prior), the paper evaluates two methods to achieve 3-D tracking:
1. Planar Array: Replacing the 1-D ULA with a 2×4 Uniform Planar Array (UPA) to measure elevation angle directly.
2. Multistatic Geometry: Adding a second nrUE transmitter to create a second bistatic pair, using range diversity to geometrically resolve altitude.

Key Contributions

End-to-End O-RAN ISAC Testbed: A live 5G NR stack integrated with a Near-RT RIC and a ray-traced channel emulator. Sensing runs inside the OAI gNB PHY, and detections reach the xApp via a prototype E2 service model, all without modifying the NR standard.
Bistatic Tracker Evaluation: An evaluation of an EKF xApp using a nonlinear measurement model, Hungarian association, and time-driven track maintenance. The evaluation includes adversarial tests such as height-prior ablation, street-canyon occlusion, and sensing under concurrent 10 Mbit/s uplink traffic.
Two Routes to 3-D Tracking: The paper demonstrates two distinct methods to remove the height prior and achieve altitude observability:
- A vertical aperture (UPA) at the gNB.
- Range diversity from a second bistatic pair.
  Both methods are shown to collapse the cross-seed spread of the final altitude estimate from ~12.7 m (unobservable) to ~0.2 m (observable).

Results

Detection Performance: In a flat line-of-sight (LOS) scenario, the system achieves 92.7% detection coverage. The measurement accuracy (vs. ray-traced ground truth) is:
- Excess bistatic range RMSE: 8.66 m (approx. one resolution cell).
- Bistatic range-rate RMSE: 0.44 m/s.
- Azimuth RMSE: 7.02°.
Tracking Performance:
- 2-D EKF (with height prior): Achieves a range (x) RMSE of 7.26 m and cross-range (y) RMSE of 5.32 m. Track continuity is 54.2% (rising to 88.3% with coast-bridging).
- 3-D EKF (Planar Array): With a 2×4 UPA, the system estimates altitude with an RMSE of 5.3 m and horizontal position with 3.8 m RMSE. The cross-seed spread of altitude drops to 0.2 m, confirming observability, though a 6.0 m bias remains.
- 3-D EKF (Multistatic): With a second transmitter, the system achieves an altitude RMSE of 2.2 m and horizontal RMSE of 5.6 m. Crucially, this configuration is unbiased (55.3 m estimate vs. 55 m truth) and also achieves a 0.2 m cross-seed spread.
Consistency: The filter exhibits structural over-confidence (NEES >> expected) due to unmodeled cross-axis correlations and the limitations of the 2-D state with a fixed height prior, but the measurements themselves are physically sound.

Significance and Claims
The paper claims to present the first end-to-end 5G O-RAN ISAC simulation testbed for UAV tracking that includes extensive pipeline validation and identifies key limitations for practical deployment. Its primary significance lies in demonstrating that:

Feasibility: Existing 5G uplink signals (UL-SRS) can be repurposed for passive UAV detection and tracking within a standard-compliant O-RAN architecture without spectrum changes.
Observability: Altitude observability, which is impossible with a single bistatic pair and a 1-D array, can be established end-to-end through the live stack using either a vertical receive aperture or a second transmitter.
Architecture: The separation of heavy signal processing (clutter cancellation, RD mapping) inside the gNB PHY and algorithmic tracking (EKF) in the Near-RT RIC xApp is a viable architectural pattern for ISAC, mirroring the emerging "sensing-dApp" model.

The authors remain modest regarding their claims, noting that the testbed is a simulation (using Sionna RT) and that the sensing service model (SM-SENS) is a working demonstration rather than a standardized specification. They identify the need for hardware integration (USRP validation) and further work on multi-target tracking and identifier consolidation as future steps.

More like this