Abstract

This thesis proposes a comprehensive perception-planning-control framework for autonomous early-stage wildfire detection and suppression using Unmanned Aerial Vehicles (UAVs).
For perception, a novel dual-stream object detection model based on the YOLOv8n architecture leverages both visible and infrared imagery after image registration. By integrating a Channel Prior Convolutional Attention (CPCA) module and a Dual Modality Cross-attention Transformer Fusion (DMCTF) module, the network effectively fuses cross-modal features while actively suppressing thermal noise and visual artifacts. Its feature extraction transparency is visually validated using Grad-CAM.
For planning, to optimize flight distance, Dynamic Programming (DP) and Simulated Annealing (SA) algorithms are implemented for single-UAV, multiple-fire-spot path planning, while a Genetic Algorithm (GA) handles multi-UAV, multiple-fire-spot cooperative path planning. For control, a Linear Quadratic Tracker (LQT) ensures precise trajectory tracking.
The framework is rigorously validated through MATLAB/Simulink and DJI Assistant 2 simulations and outdoor tests. Utilizing a DJI Matrice 300 RTK UAV equipped with an H20T camera and a customized multi-drop solenoid-based mechanism, the system successfully demonstrates autonomous wildfire detection, optimal path planning and tracking, and targeted fire retardant deployment in a single mission.