A Vision-Based Aerial-Ground Fusion Method for Fire Detection

Abstract

Wildfires are uncontrolled, destructive events characterized by rapid onset, high intensity, and significant challenges in both control and suppression. While satellite-based detection systems are limited by their resolution and inability to provide real-time data, ground-based and aerial systems offer faster response times and higher spatial accuracy, making them better suited for early wildfire detection.In this study, we propose an innovative vision-based aerial-ground fusion detection and localization framework for wildfire detection. It includes ground-based stereo vision positioning system, UAV detection, and decision-level fusion framework.To address the issue of inaccurate localization in the early stages of wildfires, particularly when the flame occupies a small area in the image, we developed a ground-based stereo vision localization system that integrates feature point matching with an object detection filtering module. Then, the algorithm is deployed on lightweight edge computing devices powered by AI.  Considering the adaptability of multi-source  data fusion from the air and ground, we designed a decision-level fusion method to integrate the detection results from both the ground and UAVs. An improved Bayesian method is used to calculate fusion confidence, enhancing the reliability of wildfire assessments. Experimental results show that the relative error in ground-based localization method under different lighting conditions and distances ranges from 3.78% to 14.89%. Compared to ground-based and UAV-based single-source detection, the aerial-ground fusion method improves the F1-score by 10.4% and 91%, respectively.

Keywords: Fire and smoke detection, Aerial-ground fusion, Stereo camera, Vision-based.