A Descriptor-less Well-Distributed Feature Matching Method Using Geometrical Constraints and Template Matching

The problem of feature matching comprises detection, description, and the preliminary matching of features. Commonly, these steps are followed by Random Sample Consensus (RANSAC) or one of its variants in order to filter the matches and find a correct model, which is usually the fundamental matrix. Unfortunately, this scheme may encounter some problems, such as mismatches of some of the features, which can be rejected later by RANSAC. Hence, important features might be discarded permanently. Another issue facing the matching scheme, especially in three-dimensional (3D) reconstruction, is the degeneracy of the fundamental matrix. In such a case, RANSAC tends to select matches that are concentrated over a particular area of the images and rejects other correct matches. This leads to a fundamental matrix that differs from the correct one, which can be obtained using the camera parameters. In this paper, these problems are tackled by providing a descriptor-less method for matching features. The proposed method utilises the geometric as well as the radiometric properties of the image pair. Starting with an initial set of roughly matched features, we can compute the homography and the fundamental matrix. These two entities are then used to find other corresponding features. Then, template matching is used to enhance the predicted locations of the correspondences. The method is a tradeoff between the number and distribution of matches, and the matching accuracy. Moreover, the number of outliers is usually small, which encourages the use of least squares to estimate the fundamental matrix, instead of RANSAC. As a result, the problem of degeneracy is targeted at the matching level, rather than at the RANSAC level. The method was tested on images taken by unmanned aerial vehicles (UAVs), with a focus on applications of 3D reconstruction, and on images taken by the camera of a smartphone for an indoor environment. The results emphasise that the proposed method is more deterministic rather than probabilistic and is also robust to the difference in orientation and scale. It also achieves a higher number of accurate and well-distributed matches compared with state-of-the-art methods