A Fast and Robust Attitude Estimation Method Based on Vision-Inertial Tight Coupling With Unknown Correspondences

Simultaneous pose and correspondence determination (SPCD) plays a crucial role in attitude estimation with unknown correspondences. In contrast to the conventional loose integration of optimizing the initial pose selection of SPCD algorithms using inertial information, we refine the data collected by the inertial sensors into motion information and tightly fuse it with the feature point information provided by the camera. As a result, the accuracy and computational speed of the attitude estimation are improved. Specifically, we propose a novel SPCD method that combines motion information-assisted feature point hybrid tracking (MIFHT) and cascaded square root cubature Kalman filter (CSRCKF), along with a vision-inertial tightly coupled framework incorporating robust strategies. Our method maintains high tracking accuracy even in visual occlusion and exhibits robustness against outliers. Finally, comparative experiments conducted on our measurement system validate the effectiveness and superiority of our method. While maintaining a convergence rate of 95%, the computational speed improves by 55.11% compared to the fastest existing SPCD method.