A robust INS/PnP fusion method for GNSS-denied navigation of small UAVs

Abstract

This thesis presents three INS/VISNAV integration architectures to deal with prob- lem when operating UAVs. The first two integration architectures are called tight and loose INS/VISNAV integration, respectively. The tight and loose integration strategies are not new and have been used before. The tight approach fuses INS information with camera measurements at pixel level. It is considered as an optimal approach in terms of estimation accuracy. However, it has a tendency to diverge under certain conditions: (1) Unfavorable relative geometry between the camera and feature points used to con- struct the VISANV solution and (2) Large errors in the position and attitude solution about which the tight integration measurement equations are linearized. This latter condition can occur when VISNAV updates are infrequent or spaced far apart in time. Maintaining stability of the filter that fuses camera and INS information can be chal- lenging when low quality (consumer/automotive grade) inertial sensors are used or the measurement update is less frequent. The loose integration approach is proposed as an alternative for solving the divergent problem. The integration is loose in the sense that integration occurs at the level of position and attitude. This is in contrast to tight integration where information fusion occurs at the pixel level. While it is sub-optimal from a filtering point of view, the loose integration approach can be more robust to linearization errors which lead to solution divergence. A method for computing the covariance of position and attitude estimates of the VISNAV solution is presented. The complementary advantages of loose and tight integration leads to the develop- ment of a novel, third sensor fusion methodology that enhances the robustness (here defined as resistance to divergence) of filters used to mechanize camera-aided inertial navigation systems (INS) while preserving the estimation accuracy. This third approach is a hybrid filter that can switch between the optimal tight and suboptimal loose strate- gies "on the fly" depending on the geometry of the landmarks being tracked and the quality of the inertial sensor. The fusion strategy is based on dual hypothesis testing approach. The proposed approach has the advantages of enhancing the robustness while maintaining the estimation accuracy.