Line Model-Based Drift Estimation Method for Indoor Monocular Localization

Currently, monocular localization as a category of vision-based indoor localization methods has attracted more attention because of its application to indoor navigation and augmented reality. In a typical monocular localization system, the absolute position estimation is utilized to acquire the initial position of the query camera, and then the relative position estimation is employed to achieve subsequent camera positions. However, accumulative errors, i.e., localization drifts, caused by the relative position estimation seriously affect the localization performance. Therefore, a line model that contains a fitted line segment and some visual features is introduced, and the line model is used to seek inliers of the estimated features for the drift estimation. Based on the pre-constructed dense 3D map, a line model-based drift estimation method is presented to monitor accumulative errors. As a switching mechanism, the proposed method determines when the relative position estimation should switch to the absolute position estimation to correct user positions. Compared with the existing monocular localization methods, the proposed drift estimation method significantly reduces the accumulative errors, and the average errors are limited within a desirable range by giving a proper drift threshold. Experimental results demonstrate that the average localization errors of the proposed method are limited within 30 centimeters in various scenes by a 50-centimeters drift threshold.