Adaptive Data Transport Mechanism for UAV Surveillance Missions in Lossy Environments

Niloufar Mehrabi, Sayed Pedram Haeri Boroujeni, Jenna Hofseth, Abolfazl Razi, Long Cheng, Manveen Kaur, Jim Martin, Rahul Amin

Introduction

Unmanned Aerial Vehicles (UAVs) have become indispensable tools for Intelligence, Surveillance, and Reconnaissance (ISR) missions. However, their operational efficiency is often hampered by a "bottleneck" trifecta: limited battery life, constrained payloads, and restricted communication bandwidth. Standard video compression techniques like H.264 or HEVC prioritize overall visual fidelity, which often results in transmitting redundant background data at the expense of mission-critical information.

System Architecture

To address these challenges, researchers have developed an AI-driven "selective transmission" strategy. This approach prioritizes sending only the most vital parts of a video frame—such as a moving vehicle—rather than the entire scene.

System architecture showing the feedback loop between the UAV and the Ground Control Station (GCS).

The proposed framework operates as a closed-loop system. The heavy computational tasks, including object detection (YOLOv8) and transmission scheduling (Deep Q-Network), are offloaded to a Ground Control Station (GCS). The GCS processes incoming data and sends a lightweight "mask" back to the UAV, instructing it on which image patches to prioritize in subsequent frames.

Selective Transmission via Deep Reinforcement Learning

The core of this mechanism is a mission-aware transport layer that treats video frames as a grid of $K \times K$ patches. A Deep Q-Network (DQN) assigns transmission probabilities to each patch based on its relevance to the mission.

Visual representation of the selective transmission: (a) Heatmap showing patch importance scores; (b) Reconstructed frame focusing on the target vehicle.

Key components of the methodology include:

  • Tiling: Dividing frames into discrete patches for granular control.
  • Predictive Intelligence: The DQN learns motion patterns to anticipate where an object will move, ensuring the system maintains a "lock" even in lossy environments.
  • Regularized Objective Function: A penalization term prevents abrupt shifts in patch selection, ensuring smooth detection performance.
  • Inter-frame Interpolation: The GCS fills in non-transmitted patches to maintain a coherent visual stream for human operators.

Experimental Evaluation

The system was rigorously tested using the AU-AIR dataset, which contains over 32,000 annotated frames of low-altitude aerial surveillance. The results demonstrate significant improvements over traditional or random selection methods:

  • High Efficiency: At a 50% transmission budget, the system maintains a 98% F1 score for object detection.
  • Robustness under Constraints: Even when data transmission is reduced by 95%, the system achieves a 91% F1 score, outperforming random selection by 87%.
  • Real-Time Capability: The mechanism supports streaming at 30 FPS with negligible processing delay.

Conclusion and Future Work

This research shifts the focus from general visual fidelity to "mission success." By leveraging existing low-bandwidth feedback channels, the system provides a scalable solution for bandwidth-constrained UAV operations. Future work will aim to extend this framework from Single Object Tracking (SOT) to complex Multi-Object Tracking (MOT) scenarios in dense urban environments.

Reference: Mehrabi, N., et al. "Adaptive Data Transport Mechanism for UAV Surveillance Missions in Lossy Environments." arXiv:2410.10843v1 [eess.IV], Sep 2024.

Summary updated

BibTeX
@inproceedings{mehrabi2025adaptive,
  author = {Niloufar Mehrabi and Sayed Pedram Haeri Boroujeni and Jenna Hofseth and Abolfazl Razi and Long Cheng and Manveen Kaur and Jim Martin and Rahul Amin},
  title = {{Adaptive Data Transport Mechanism for UAV Surveillance Missions in Lossy Environments}},
  booktitle = {CCNC},
  year = {2025},
  doi = {10.1109/ccnc54725.2025.10975947},
}
Figure