基于运动平均的无地图视觉定位方法OA

To address the issue of time-consuming pre-constructed point clouds and their susceptibility to reconstruction errors in structure-based visual localization methods,a map-free visual localization method is proposed.This method directly predicts the pose from the matching pairs of query images and database images and fuses and optimizes multiple predicted poses using motion averaging.Firstly,the fundamental matrix is iteratively estimated by weighted matching points combined with epipolar geometry error.Then,the predicted pose is calculated based on the fundamental matrix and the pose of the database image,and the rotation prediction is averaged using an L1 and L2 hybrid optimization algorithm in the Lie algebra space.Finally,on the inliers with high rotational consistency,the translation prediction is optimized by minimizing the camera center angular error.Experimental results show that this method achieves accuracy comparable to mainstream methods and performs best in indoor scenes.In the practical BishanHall scene,the average rotation and repro-jection error is 0.02° and 1.172 pixel higher than that of the compared method with the same configuration,the runtime is shortened by about 0.442 s,and the storage space is reduced by about 38％by eliminating the pre-constructed point cloud map.