Observability-Informed Measurement Validation for Visual-Inertial Navigation
Type of DegreeMaster's Thesis
MetadataShow full item record
This thesis presents a measurement validation method for visual-inertial navigation and an implementation of a visual-inertial estimator which makes use of it. Many autonomous plat- forms, especially flying ones, rely on accurate and reliable state estimates from a visual-inertial estimator to maintain safe and controlled flight towards a goal. The measurement validation method presented in this thesis makes use of a geometric analysis of the landmark measuremnt model to enable early and reliable validation, safely integrating high-quality measurements into the state estimation process. First, the IMU and camera sensors are detailed along with the sensor processing techniques necessary to make use of them. Next, a detailed description of two standard visual-inertial estimation approaches is presented to develop necessary back- ground knowledge. Following these descriptions, an analysis of the relationships between the geometry of landmark observation and the accuracy and reliability of landmark estimates is per- formed, concluding with the proposal of a new validation method which delays measurement processing until the landmark is predicted to be observable. Lastly, a visual-inertial estimator is developed which makes use of the proposed method and tested on the EUROC dataset, the most common visual-inertial dataset, against several state-of-the-art estimators. In this comparison, the proposed estimator demonstrates competitive performace, reliably producing positioning errors of less than 0.5 meters over flights up to 120 meters long. Overall, the proposed method is demonstrated to be reliable and accurate in competition with significantly more advanced and complicated estimators.