Handling loop scenarios for vision-aided aerial navigation based on three-view geometry

This paper presents a new method for navigation aiding based on three-view geometry. An airborne platform is assumed to be equipped with a standard inertial navigation system and a gimbaled camera, which captures images along the flight. The images are stored and associated with a partial data taken from the navigation system. No additional sensors or a-priori information are assumed. Given a set of three overlapping images, and the associated navigation data, we develop constraints that relate the platform motion between the time instants of the three images. The scale ambiguity, embedded with pure vision-based motion estimation techniques, is resolved by incorporating navigation data attached to each image. The developed constraints are fused with an inertial navigation system using an Implicit Extended Kaiman Filter, allowing to reduce the position error, and errors in some other parameters, to the levels present when the first image was captured. Reduced computational resources are required, in comparison with other existing methods such as bundle adjustment and SLAM. The proposed method was examined for handling loop scenarios in indoor and outdoor navigation based on real images, allowing to maintain constant levels of navigation errors over time.