Online updating of synthetic vision system databases

In aviation, synthetic vision systems render artificial views of the world (using a database of the world and pose information) to support navigation and situational awareness in low visibility conditions. The database needs to be periodically updated to ensure its consistency with reality, since it reflects at best a nominal state of the environment. This thesis presents an approach for automatically updating the geometry of synthetic vision system databases and 3D models in general. The approach is novel in that it profits from all of the available prior information: intrinsic/extrinsic camera parameters and geometry of the world. Geometric inconsistencies (or anomalies) between the model and reality are quickly localized; this localization serves to significantly reduce the complexity of the updating problem. Given a geometric model of the world, a sample image and known camera motion, a predicted image can be generated based on a differential approach. Model locations where predictions do not match observations are assumed to be incorrect. The updating is then cast as an optimization problem where differences between observations and predictions are minimized. To cope with system uncertainties, a mechanism that automatically infers their impact on prediction validity is derived. This method not only renders the anomaly detection process robust but also prevents the overfitting of the data. The updating framework is examined at first using synthetic data and further tested in both a laboratory environment and using a helicopter in flight. Experimental results show that the algorithm is effective and robust across different operating conditions.