Abstract
Previous visual localization started from point correspondences (PCs) to estimate poses. This article takes the camera position as the entry point and finds that the camera position solution set rotates around an axis connected by two observed 3D points to form a surface called a torus-like surface (TLS). The relevant parameters of TLS are calculated based on PCs and camera intrinsic parameters. In order to reduce the number of solutions, this article uses four PCs to construct three TLSs. By utilizing four PCs, the pose determination problem is reformulated as the task of finding the optimal intersection of three surfaces. By using the step-size adaptive tracking method, the candidate set of intersections can be quickly and accurately found. Combining the feature information of intersections on TLS and the camera intrinsic parameters, the optimal position is obtained. Based on this position, the rotation matrix can be determined. In the synthetic data experiments and the dataset experiments based on image localization, it is shown that the visual localization based on TLS is more accurate than current state-of-the-art methods, which provides a new entry angle and effective ideas for visual localization. Its accuracy and practicality are fully demonstrated in the application test of augmented reality indoor navigation.