Appearence-based Visual Localization for Indoor Navigation of Quadrotors

The main idea behind this work is to perform mapping and localization for a quadrotor in an indoor environment topologically represented by a set of ordered key images acquired in a prior learning phase. The mapping and localization are based on a feature matching between the most recently acquired image and the set of key images. The overall task is composed of two consecutive phases: a mapping phase and a navigation phase. The mapping phase is an offline learning phase during which the quadrotor is manually moved around the environment by means of a joystick. The drone acquires a sequence of images that will represent the path the quadrotor should follow in the next phase. The navigation phase is accomplished online and it’s composed by an initial localization, a successive localization and the motion control. The quadrotor can start from any initial position in the environment and the initial localization algorithm determines the initial pose of the quadrotor by comparing the first acquired image with all the key images from the prior phase and selecting the two images of the path most similar to the acquired image. The successive localization algorithm is used, during the motion of the quadrotor, to decide when to switch the reference key images. This step is necessary in order to determine the position of the quadrotor during the motion. Finally, the control law is built to minimize the difference between the desired feature value and the current feature value, in order to keep the quadrotor on the defined trajectory.