Registration and Geolocation of Aerial Thermal Images for Wildfire Monitoring

The registration problem of thermal images captured from an Unmanned Aerial Vehicle (UAV) is addressed in this paper. After an exhaustive literature review, we conclude that our work is the first to demonstrate that the Scale Invariant Image Transform (SIFT) along with Random Sample Consensus (RANSAC) can successfully be used on thermal images. We contribute an open-source thermal-image registration system that enables UAVs to geolocate hotspots of wildfires. UAVs' measurements of geolocation and cardinal direction have errors, so the location and direction of a hotspot have different values when seen from different UAVs. This paper estimates the geolocation of hotspots by identifying SIFT keypoints on overlapping thermal images, matching the keypoints, and applying a transformation using RANSAC to reconcile the differences among UAVs. We propose image-processing operations including blurring, closing, and opening to enable the successful use of SIFT and RANSAC. Our thermal-image registration system is validated on three datasets. Images of emulated fires were collected using our open-source UAV platform for the first dataset. The other datasets are FLAME and FLAME3, which include images of controlled fires. Our results demonstrate registration and geolocation of hotspots with a success rate of 97%, 90%, and 92% respectively over a total 215 pairs of images.