Minimal solver for relative pose estimation under planar motion
Visual-based relative pose estimation is a core technology for autonomous localization and environment perception of mobile robots such as autonomous driving vehicles.In order to improve the accuracy,efficiency,and robustness of the algorithm,the minimal solvers for relative pose estimation are an important research topic.Traditional relative pose estimation algorithms typically utilize only the image coordinate information of the matched feature points,which ignores the additional information provided by feature descriptors,such as feature rotation angles and scales.In this paper,focusing on the common scenario in mobile robot applications,we propose a new minimal solver for planar motion estimation using the constraints provided by the feature descriptors.By exploiting the homography constraints,we obtain a closed-form solution for the relative pose of the monocular camera utilizing a single rotation-invariant feature.Due to the minimal number of features required by the proposed solver,it can be efficiently combined with RANSAC or histogram voting methods for initial motion estimation and removal of outlier matches.Experiments on synthetic data and public datasets demonstrate that our method significantly improves the accuracy and robustness of monocular camera planar motion estimation,which makes it applicable to autonomous localization and visual perception of mobile robots.
万方数据
