Stable Multi-Drone GNSS Tracking System for Marine Robots

This paper presents a stable, real-time multi-drone GNSS tracking system for marine robots that integrates visual detection, multi-object tracking, triangulation, and a confidence-weighted Extended Kalman Filter with cross-drone ID alignment to overcome the limitations of underwater signal loss and traditional alternatives.

The Gist

Imagine you are trying to find a friend swimming in a vast, murky lake. You can't see them underwater, and if you try to shout their location using a radio (like a GPS signal), the water swallows the signal instantly. This is the biggest headache for marine robots: they lose their "map" the moment they dip below the surface.

Traditionally, engineers have tried to solve this by giving the robot a "dead reckoning" compass (guessing where it is based on speed) or by building expensive underwater lighthouses. But these methods are either inaccurate, expensive, or require heavy infrastructure.

This paper introduces a clever, low-cost solution: A team of drones acting as a "flying GPS" for the robots.

Here is the breakdown of how it works, using simple analogies:

1. The Problem: The "Underwater Blackout"

Think of a marine robot like a submarine. When it's on the surface, it can see the sky and get a GPS signal. But the second it dives, the water blocks the signal. It's like walking into a thick fog bank; you know where you were, but you don't know where you are anymore. If the robot drifts off course, it might crash or get lost.

2. The Solution: The "Drone Swarm"

Instead of relying on one drone, the researchers use three drones flying in formation above the water.

• The Analogy: Imagine three friends standing on a hill looking down at a person walking in a field. If one friend looks away or gets blocked by a tree, the other two can still see the person.
• Why it matters: If one drone loses sight of the robot (due to a wave or a cloud), the others pick up the slack. This creates a "safety net" so the robot is never truly lost.

3. How They Find the Robot: "Triangulation with a Camera"

The drones don't have special underwater sensors. They just have cameras.

• The Process:
  1. The drone spots the robot in its camera view.
  2. It knows exactly how high it is flying and which way it is facing.
  3. It calculates the angle to the robot.
  4. By combining the angles from multiple drones, the system draws invisible lines in the air. Where those lines cross is exactly where the robot is.
• The Metaphor: It's like playing "Hot and Cold." If one drone says, "It's 100 meters to my left," and another says, "It's 50 meters to my right," the computer can pinpoint the exact spot where those two descriptions meet.

4. The "Brain": Keeping Track of IDs

This is the trickiest part. If you have three drones watching two robots, how does the computer know that "Robot A" seen by Drone 1 is the same "Robot A" seen by Drone 2?

• The Challenge: The water is choppy, the drones shake, and the robots look very similar. A standard computer might get confused and think Robot A turned into Robot B for a split second.
• The Fix (Hybrid Matching): The system uses a "double-check" method.
  • Eye Test: It looks at the picture (does the robot look the same?).
  • GPS Test: It checks the math (is the robot in the right place?).
  • The Result: Even if the camera gets shaky, the GPS math keeps the identity straight. The paper shows that without this "double-check," the system gets confused and swaps IDs constantly. With it, the tracking is rock solid.

5. The "Filter": Smoothing the Jitters

Even with good math, the data is a little noisy (like a shaky video).

• The Solution: They use an Extended Kalman Filter (EKF).
• The Analogy: Imagine watching a runner on a bumpy track. Your eyes see them jump up and down. The Kalman Filter is like a smart observer who knows, "Humans don't actually bounce that high; they are probably running smoothly, and the camera is just shaking." It smooths out the wiggles to give you a clean, accurate path.

The Results: How Good Is It?

The team tested this in real lakes with real robots.

• Accuracy: In the best conditions, they were within less than 1 meter of the robot's true location. Even in tough conditions (sharp turns, lots of wind), they stayed within 1.7 meters.
• Speed: The system runs fast enough to update the robot's location 5 times every second.
• Cost: A single onboard GPS system for a robot costs about the same as four drones. But the drone system can track multiple robots at once and doesn't require the robot to carry heavy equipment.

Why This Matters

This isn't just about finding a robot; it's about safety and efficiency.

• Search and Rescue: If a person falls off a boat, drones can track them in the water without the person needing a special radio.
• Ocean Exploration: Scientists can send a fleet of cheap robots to map the ocean floor, knowing exactly where they are without building expensive underwater cables.

In a nutshell: This paper teaches us how to use a team of flying eyes (drones) and a smart brain (algorithms) to keep a "blind" robot (underwater) from getting lost, turning a chaotic ocean into a navigable map.

Technical Summary

Here is a detailed technical summary of the paper "Stable Multi-Drone GNSS Tracking System for Marine Robots."

1. Problem Statement

Global Navigation Satellite System (GNSS) signals are unavailable for marine robots once they submerge even slightly below the surface or if their antennas are wet. Traditional onboard localization methods (e.g., Inertial Measurement Units combined with Doppler Velocity Logs) suffer from error accumulation (drift), while acoustic methods (LBL/USBL) require expensive fixed infrastructure. Existing offboard solutions using a single drone are prone to occlusion, limited coverage, and single points of failure. There is a critical need for a robust, scalable, and low-cost system to provide continuous, accurate localization for surface and near-surface marine robots without relying on underwater infrastructure.

2. Methodology

The proposed system utilizes a multi-drone aerial platform to visually track marine robots and estimate their GNSS coordinates in real-time. The pipeline consists of four main stages:

• Data Acquisition & Preprocessing:

  • Field trials were conducted in freshwater using marine robots equipped with ground-truth GNSS receivers.
  • Three drones captured video from varying altitudes and angles.
  • Data augmentation (rotation, motion blur, glass blur for water distortion) was applied to train the vision model for robustness against environmental variations.

• Visual Detection and Tracking:

  • Detection: A fine-tuned YOLO v11 model detects marine robots in drone video streams.
  • Tracking: The ByteTrack algorithm is used for multi-object tracking due to its computational efficiency and ability to handle uncertainty. Each drone runs tracking independently.

• GNSS Position Estimation (Triangulation):

  • The system calculates the robot's 3D position relative to the drone using camera geometry.
  • It converts pixel coordinates to angular displacements ($\theta_x, \theta_y$) using focal length and sensor dimensions.
  • Using the drone's altitude ($A$) and camera tilt ($\alpha$), it computes the ground distance ($D$) and radial distance ($D_r$).
  • These local coordinates are transformed into global GNSS latitude and longitude using the drone's heading ($\psi$) and conversion factors.

• State Estimation and ID Alignment:

  • Confidence-Weighted EKF: An Extended Kalman Filter (EKF) fuses position estimates from multiple drones. The fusion is weighted by the detection confidence scores of the YOLO model to prioritize high-quality observations.
  • Hybrid Matching: To prevent ID switches caused by camera shake or occlusion, the system computes a similarity score combining Intersection over Union (IoU) (image space) and Haversine distance (GNSS space). The weights are set to $w_1=0.7$ (IoU) and $w_2=0.3$ (GNSS).
  • Cross-Drone ID Alignment: A centralized algorithm aligns tracking IDs across different drones. One drone acts as a reference; the Hungarian algorithm is used to match detections from other drones to reference tracks based on GNSS distance, ensuring global consistency.

3. Key Contributions

• Multi-Drone Redundancy: The system employs multiple drones to ensure overlapping coverage, mitigating the risks of occlusion and tracking loss inherent in single-drone setups.
• Hybrid Matching Strategy: A novel matching metric that fuses visual IoU with GNSS proximity significantly reduces ID switching, particularly in turbulent conditions where visual tracking alone fails.
• Global ID Consistency: An algorithm to synchronize tracking IDs across distributed drone views, enabling robust multi-robot tracking in a shared coordinate frame.
• Low-Cost, Infrastructure-Free Solution: The approach eliminates the need for underwater beacons or expensive onboard sensors, relying instead on standard drones and vision processing.

4. Results

The system was evaluated across five test categories involving linear and non-linear trajectories, varying numbers of drones (1, 2, and 3), and challenging maneuvers (sharp U-turns).

• Accuracy:
  • 3 Drones (Linear): Mean error 0.942 m (STD 0.743 m).
  • 3 Drones (Non-Linear): Mean error 1.524 m (STD 1.435 m).
  • 3 Drones (Hard Turns): Mean error 1.732 m (STD 1.481 m).
  • Even with a single drone, the system maintained sub-1.2m accuracy, though performance degraded slightly with fewer drones.
• ID Stability: Under strong air turbulence, the hybrid matching method achieved zero ID switches, whereas a standard IoU-only approach averaged 1.33 switches per 500m.
• Real-Time Performance: Deployed on an NVIDIA Jetson Xavier, the system processes the full pipeline in ~0.176 seconds per frame, achieving a 5 Hz update rate, sufficient for real-time control.

5. Significance

This work demonstrates a practical, scalable solution for marine robotics localization that bridges the gap between satellite coverage and underwater operations.

• Operational Impact: It enables persistent tracking of multiple robots (or even non-instrumented objects like humans in search and rescue) without requiring the robots to surface frequently for GPS recalibration or carry heavy acoustic equipment.
• Scalability: The distributed nature of the system allows for easy expansion to larger drone fleets, potentially improving accuracy further.
• Cost-Effectiveness: The total cost of the multi-drone setup is comparable to a single high-end onboard localization system, yet it offers superior capabilities for multi-robot coordination and surface-level monitoring.
• Future Potential: The framework lays the groundwork for broader applications in marine research, environmental monitoring, and conservation (e.g., tracking schools of fish), with future work planned to explore particle filters and decentralized voting mechanisms for even greater robustness.