Global localization of nano drone in an indoor environment

Abstract: For a drone to be able to navigate in an indoor environment, it needs to understand its surroundings and locate itself to be able to plan a trajectory to its final destination. This thesis aims to solve the global localization problem, i.e. estimate a drone’s position and orientation in a previously mapped indoor environment by using a monocular camera and computer vision, which is an important first step towards autonomous navigation To make a drone able to understand its surroundings, a camera is attached to it and computer vision algorithms are used to extract important information about features in the environment represented in images. Using an open source software, COLMAP the features can be recreated in a map. A feature in the map is represented by a 3D-point and a descriptor, which describe the location and the structure of the feature in the world. Many points create together a point cloud. To be able to use the point cloud as a map for navigation, the scale ambiguity problem needs to be solved. Because of similarity properties of the projection model used in COLMAP, the point cloud can have arbitrary orientation and scale. A distance in the map can then be arbitrarily big, which makes it impossible to plan a trajectory. Therefore, the point cloud is rotated to match the orientation of the gravity direction and is scaled to metric scale by using sensor data from i.a. the drone’s IMU. By extracting features from images when the drone is flying, descriptors representing the features can be computed and compared with descriptors in the point cloud map. Point correspondences are then generated between the map and image. They are later used to solve the Perspective-three-point problem to derive a pose estimate of the drone in the environment.