Multi-Drone GNSS Tracking

• The topic presents a multi-drone GNSS tracking system that integrates distributed sensor fusion and robust consensus algorithms to ensure resilient localization even with intermittent GNSS signals.
• Methodologies employ Kalman filters, vision-based triangulation, and extended Kalman filters to enhance state estimation and target tracking accuracy in hybrid operational settings.
• Key applications include UAV swarm coordination and marine target tracking, with performance validated by improvements in RMSE, consensus convergence times, and reduced ID-switch rates.

A multi-drone GNSS-based tracking system refers to a coordinated collection of unmanned aerial vehicles (UAVs) equipped with Global Navigation Satellite System (GNSS) receivers and supplementary sensors for robust, scalable localization and target tracking, especially in environments with partial or intermittent GNSS availability. These systems leverage sensor fusion, distributed consensus, and communication protocols to provide reliable estimates of either the drones’ own states or those of tracked ground/marine targets. Key paradigms fall into two broad categories: (1) distributed drone self-localization/resilience (notably SwarmRaft (Dev et al., 1 Aug 2025)), and (2) multi-drone cooperative tracking of external targets, such as marine robots, by aerial platforms (notably (Wen et al., 24 Nov 2025, Wen et al., 7 May 2025, Mason et al., 2020)). Approaches employ Kalman filters, bearing triangulation via vision, and consensus mechanisms to achieve high robustness and accuracy in GNSS-challenged settings.

1. System Architectures and Sensing Modalities

The hardware and network architecture of multi-drone GNSS-based tracking systems consists of a constellation of UAVs, each with a GNSS receiver, inertial navigation sensors (IMU/INS), and wireless communication modules. Some systems incorporate additional on-board components:

• For drone self-localization and swarm coordination (Dev et al., 1 Aug 2025): Each UAV carries GNSS, IMU, UWB/radio-based inter-drone ranging, and dedicated modules running Kalman filter-based state estimation and a Raft consensus agent.
• For external target tracking (marine robots, vehicles) (Wen et al., 24 Nov 2025, Wen et al., 7 May 2025): Drones combine downward-facing calibrated RGB cameras, altimeters, IMU, and integrated compute units for real-time vision processing, with communication via Wi-Fi or mesh networks.

System topologies vary depending on application: SwarmRaft employs a fully connected peer-to-peer exchange, while marine tracking architectures support mesh or star topologies, with ground stations aggregating results in some scenarios (Wen et al., 24 Nov 2025, Wen et al., 7 May 2025).

Table: Typical Hardware Components

System Type	GNSS	IMU/INS	Camera	Ranging	Communication	Compute
Swarm Coordination	✓	✓	(vision opt.)	✓ (UWB/RSSI)	Wi-Fi broadcast	Onboard
Marine Tracking	✓	(✓)*	✓	—	Wi-Fi/mesh	Onboard
LoRaWAN Tracking	✓	✓	—	—	LoRaWAN	Minimal

*Optional: for attitude/height estimation

2. State Estimation, Sensor Fusion, and Consensus

Drone Self-Localization (SwarmRaft Paradigm):

Drones maintain an augmented state vector

$$\mathbf s_{i,k} = \begin{bmatrix} \mathbf x_{i,k} \ \mathbf v_{i,k} \ \psi_{i,k} \end{bmatrix} \in \mathbb{R}^7$$

for position, velocity, and heading at time step $k$. State propagation and correction use discrete-time Kalman filtering, integrating GNSS when available and falling back to IMU/INS prediction otherwise. Fault detection and correction are mediated by majority voting and median-based peer fusion, under Raft-style consensus to mitigate GNSS spoofing/failure and sensor faults (Dev et al., 1 Aug 2025).

Cooperative Target Tracking:

Multi-drone tracking of marine robots relies on geometric camera-based triangulation of detected targets, broadcasting per-drone GNSS pose and image observations. Each detection is projected to the global frame using both camera geometry and GNSS, then fused via an Extended Kalman Filter (EKF) to attenuate error (Wen et al., 24 Nov 2025, Wen et al., 7 May 2025):

$$\mathbf{x}_k = [X_k, Y_k, \dot{X}_k, \dot{Y}_k]^\top,\quad \mathbf{x}_{k|k-1} = F\,\mathbf{x}_{k-1|k-1},\quad P_{k|k-1} = F\,P_{k-1|k-1}\,F^\top + Q$$

Measurements (from bearing triangulation or direct GNSS) are assimilated into the filter to yield world-frame estimates with real-time update rates.

Communication-Efficient Tracking (LoRaWAN):

UAVs periodically report compressed state vectors (position, velocity, heading, turn rate, acceleration) via LoRaWAN uplink packets. The central station applies a 7-dimensional EKF or UKF, using an extended 3D Constant Turn Rate and Acceleration (CTRA) motion model to predict and update UAV tracks (Mason et al., 2020).

3. Multi-Drone Coordination, Assignment, and Data Association

Assignment and coordination strategies are needed to ensure system scalability and target handoff:

• Leader election and consensus: SwarmRaft leverages the Raft protocol for leader election and log replication in a synchronized round-based regime, completing consensus within one cycle to maintain cohesive multi-drone state updates even under communication loss or GNSS attacks (Dev et al., 1 Aug 2025).
• Dynamic region assignment: Marine tracking systems partition the area into Voronoi-like regions or periodically solve a linear assignment problem (Hungarian algorithm) to allocate robots to the most suitable drone, including handoff protocols for drones entering/leaving coverage regions (Wen et al., 24 Nov 2025, Wen et al., 7 May 2025).
• Data association and global ID alignment: To achieve robust target identity coherence across drones, cross-drone assignment algorithms align local track IDs to a global reference master using Haversine-based cost matrices, Hungarian matching, and confidence thresholds (Wen et al., 24 Nov 2025).

4. Communication Protocols and Scalability Considerations

SwarmRaft and vision-based multi-drone marine trackers rely on Wi-Fi (IEEE 802.11 n/ac) networks for authenticated, broadcast, or mesh communication (Dev et al., 1 Aug 2025, Wen et al., 24 Nov 2025, Wen et al., 7 May 2025). Message sizes and frequency are optimized for real-time operation with minimal overhead:

• SwarmRaft: ~40 B per log entry, leader $O(n)$ messages per round, followers $O(1)$, with update frequency set to 5 Hz to match GNSS (Dev et al., 1 Aug 2025).
• Vision trackers: 5 Hz broadcast of pose/detection data; peer-to-peer fusion supports scale-out by adding drones (Wen et al., 24 Nov 2025, Wen et al., 7 May 2025).
• LoRaWAN-based systems: Duty-cycle constraints necessitate low-bitrate (13 B/pkt), favoring sparse periodic reporting with appropriate spreading factor and channelization to mitigate collisions (Mason et al., 2020).

Scalability is validated up to 15 drones in SwarmRaft simulations (Dev et al., 1 Aug 2025) and for up to 3–4 marine robots in vision-based systems, with guidance for augmenting coverage by adding drones.

5. Error Analysis, Robustness, and Performance Evaluation

Performance is quantified using RMSE/MAE for world-coordinate estimates, consensus convergence time, update rates, and ID-switch rates:

• SwarmRaft (Dev et al., 1 Aug 2025): For $N=\{5,10,15\}$ with up to 5 spoofed/faulty nodes, SwarmRaft achieves MAE ≈ 2 m (vs. degraded GNSS MAE up to 13.4 m). Consensus converges in ≈ 150–180 ms (< cycle time), with message loads ~5 kB/s per node.
• Multi-drone vision trackers (Wen et al., 24 Nov 2025): RMSE is in the 0.94–1.73 m range for 1–3 drones and various trajectory complexities; standard deviation increases with path complexity. ID-switches per 500 m are nearly eliminated using hybrid matching, compared to IOU-only baselines.
• Aerial marine localization (Wen et al., 7 May 2025): Mean error improves from 3.1 m (single drone, 3 robots) to projected 2.2 m with 2 drones (more coverage); onboard update rates ~4 Hz.
• LoRaWAN trackers (Mason et al., 2020): Achieve RMSE_x,y ≈ 0.9 m, RMSE_z ≈ 2 m, with update intervals 2–10 s; performance is bounded by packet loss, duty-cycle, and motion model fidelity.

Dominant error sources include GNSS noise, altimeter error, pixel detection error, and, in marine setups, the geometric projection from camera to ground coordinates and attitude estimation uncertainty.

6. Extensions, Best Practices, and Applications

• Extensions: System accuracy and robustness can be improved by integrating vision-based pose estimates, thermal/IR imaging for low-light, barometric-INS fusion for rough seas, and batch fusion or covariance intersection schemes for inconsistent multi-drone tracks (Wen et al., 24 Nov 2025, Wen et al., 7 May 2025).
• Operational best practices: Calibration (intrinsic/extrinsic) and in-flight geometric correction using known landmarks are essential for consistent performance. Data augmentation during training and adaptive filtering help maintain reliability in variable environmental conditions.
• Applications: These architectures have been validated in UAV swarm coordination (Dev et al., 1 Aug 2025), surface and near-surface marine robot localization (Wen et al., 24 Nov 2025, Wen et al., 7 May 2025), and remote management of UAV fleets via low-rate communication links (Mason et al., 2020).

The multi-drone GNSS-based tracking paradigm forms the basis for resilient, scalable, and real-time localization in large-scale, infrastructure-free deployments for environmental monitoring, marine robotics, and autonomous aerial swarms.