Real-Time Georeferencing of Fire Front Aerial Images Using Structure from Motion and Iterative Closest Point

Abstract

This work proposes the use of Structure-from-motion (Sfm) and Iterative Closest Point (ICP) as a forest fire georeferencing algorithm to be used with footage captured by an aerial vehicle. Sfm+ICP uses the real time video captured by an aircraft’s camera, as well as its Inertial Measurement Unit (IMU) and Global Positioning System (GPS) measurements, to reconstruct a dense three dimensional (3D) point cloud of the disaster area. The Sfm reconstruction is divided in two steps to improve computational efficiency: a sparse reconstruction step using Speeded-up robust features (SURF) for camera pose estimation, and a dense reconstruction step relying on a Kanade-Lucas-Tomasi (KLT) feature tracker initialized using the minimum eigenvalue algorithm. In addition, the dense 3D reconstruction is registered to a real Digital Elevation Model (DEM) of the surrounding area, and used as the basis of the georeferencing estimates. Indeed, the algorithm was validated with a real forest fire video and compares favourably with a direct georeferencing method evaluated in the same scenario. The results demonstrate that Sfm+ICP can perform accurate 3D reconstructions while also real-time georeferencing several targets in a forest fire scenario. Furthermore, the algorithm is robust to high IMU and GPS errors, making it a far better option than optic-ray-based georeferencing for UAVs with unreliable telemetry.