On the selection of candidates for point and line correspondences

The paper addresses the selection of possible candidates for point or line correspondences between model and image features in the context of pose estimation or navigation tasks. The selection incorporates a priori knowledge about the world position of the vehicle carrying the camera as well as the uncertainty of this position, the uncertainty of the camera position relative to the vehicle, spatial uncertainties of the model, and the uncertainty of the internal camera parameters, all given by covariance matrices. Using this a priori knowledge leads to a significant reduction of the number of possibly corresponding image features for each model feature without loss of correct candidates. Thus, the robustness and efficiency of the subsequent matching between model and image is increased. The presented framework replaces a widespread heuristic to achieve such a reduction by defining search spaces of fixed size and shape around each projected model feature and choosing some thresholds for the maximum deviation of properties like the 2D orientation in case of model lines