Visual odometry with failure detection for the aegis UAV

In this paper, a visual odometry system that has been developed to help solve the navigation problem on a UAV is presented. This system is part of the vision-based positioning system that will be used on a new UAV currently in development at Cranfield University. Using images captured from a single camera, the ego-motion of the aircraft is estimated and the relative position is updated every time a new frame is processed. To understand how this is achieved, each of the steps in the implemented visual odometry algorithm is explained in detail, taking a look to the techniques that make it possible. In addition, a failure detection system based on the corner tracking error has also been added to the algorithm to make the system more robust and able to automatically deactivate in poor conditions. Following the description of the visual odometry system, its performance is evaluated using terrain images from Google Earth (GE) and also from a real aerial footage captured on board a Curtis Pitts aircraft. The influence on the error of several factors such as the altitude, the flight speed, the terrain type and the number of corners to be tracked is studied and explained in detail. Finally, the future work to integrate the visual odometry system with a geolocation system to produce the vision-based positioning system of the Aegis UAV is briefly outlined.