Real-Time Multimodal Fusion and Explainable Decision-Making for Autonomous Reversible Counter-UAV Systems

Sammanfattning

The rapid proliferation of Unmanned Aerial Vehicles (UAVs) has increased privacy, safety, and security risks in urban airspace. However, existing Counter-UAV (C-UAV) systems are often infrastructure-heavy, costly, and primarily designed for large-scale deployments, which limits their suitability for resource-constrained and civilian-oriented use. To address this gap, this thesis investigates how a real-time multimodal fusion and explainable decision-making pipeline can be designed for autonomous reversible C-UAV systems under resource constraints. Following a Soft Design Science Research approach, this study designs and evaluates a modular ROS2-based artifact consisting of a lightweight vision detection branch, a lightweight acoustic detection branch, a confidence-weighted late-fusion decision layer with interpretable state-based logic, and a reversible GPS-spoofing-based action mechanism implemented in simulation. The artifact is evaluated through scenario-based experiments covering modality-level verification, multimodal fusion, decision-state logic, mitigation triggering, and system-level runtime performance in a ROS2–Gazebo–PX4 environment. The results show that the visual and acoustic branches produce stable outputs under controlled positive and negative conditions, the fusion and decision layer generate graded and operationally explainable threat states, and the action layer can be activated, maintained, stopped, and re-armed in simulation. System-level observations further indicate that the acoustic and decision branches remain computationally lightweight, while the main runtime bottleneck lies in the YOLO-based visual detector. Standalone Raspberry Pi 5 tests further confirm deployment feasibility on edge-class hardware, although with limited frame rate. These findings suggest that a coherent multimodal, explainable, and reversible C-UAV pipeline can be realized under resource-constrained conditions. At the same time, the current artifact should be understood as a simulation-oriented proof of concept rather than a fully field-validated deployment solution.