In recent times, drones are used in a wide range of applications. Many of these, such as wildfire or flood monitoring, include the classification of certain areas. Delivering drone images with these specific areas highlighted can provide a lot of information but leaves the association between the 2D image and the 3d world to the user. It would therefore be preferable to have a method that connects the information gained from the drone images to 3d positions.
The authors were confronted with this challenge during their internship at Robotto, where they developed an algorithm to autonomously detect and map wildfire. During this internship, a simple projection from the drone image to an assumed ground plane was used to map the found fires. The accuracy of this method proved to be unsatisfactory. The aim of this master thesis is therefore to develop an improved method. For this purpose, an algorithm was developed that characterizes an area which is visible in many subsequent drone images as a 3d point cloud. It accomplishes this by tracking landmarks around the border of the target area and using Kalman filtering to find the 3d position of the individual landmarks.
This algorithm was tested using state of the art simulation and showed promising results. It digests images subsequently and starts providing results as soon as possible. This means that, once optimized, it could run on the fly and provide input to the path planning of a drone. This differentiates it from alternatives, such as photogrammetry, that require the whole dataset to start processing.