A Log-Polar Feature Guided Iterative Closest Point Method for Image Registration

Images can substantially change their appearance and shape when they are acquired using different modalities, with lighting variations or at widely different viewpoints. Even with the state-of-the-art technology, e.g., the generalized dual-bootstrap iterative closest point (GDB-ICP) method, it is still difficult to register those challenging images. To handle this issue, this paper proposes a novel log-polar feature guided iterative closest point (LPF-ICP) method for image registration. In particular, by taking log-polar (LP) image features (including corners and bulbs) as seed, the LPF-ICP method first forms the initial bootstrap regions and the related similar transformation via the matching of LP seeds. Then, driven by the single-scale Harris features (including corners and edges), the proposed method gradually expands the bootstrap regions and updates the transformation estimate until the regions cover the entire image overlap. Finally, the experimental evaluation shows that the LPF-ICP method succeeds in registering all the 22 image pairs contained in the Rensselaer dataset, while the GDB-ICP method only succeeds 19 of them.