Vision based landmark detection for UAV navigation

The majority of Unmanned Aerial Vehicles (UAV) available today depend on Global Position Satellites (GPS) and inertial measurement units (IMU) for state estimation used in navigation and control. However with the increase in availability of cheap GPS jamming technologies leads to concerns over the dependence of GPS for control and navigation. A possible solution is to use a downward looking camera on-board the aircraft, and using vision based techniques the aircraft can estimate its position without the need for GPS signals. The focus of this thesis is to develop reliable methods for feature and landmark extraction for use with the vision based positioning system. The first method proposed estimated the aircraft position in real time using Image Registration techniques, during testing it was found that it did not cope well if there are differences between the source and reference images, which could be due to seasonal or lighting changes. To overcome this problem, work was conducted to look at object detection (buildings, and roads) which enable objects to be detected despite changes in season, or lighting conditions. Three such methods are presented in this thesis, although all of them have been shown to work, only the Haar classifier based method is suitable for use on-board a UAV as the other methods are computationally intensive. Further testing of the Haar classifier was conducted to investigate the full envelope of the object detection under a simulated test. Haar classifier cascade for object detection in aerial images was shown to be capable of detecting objects reliably under a variety of different situations in this thesis. This information can then be used with a GIS database to match the objects extracted from the image, with objects on a geo coded object database to estimate the aircraft position in a variety of different conditions.