Orthogonal corner detection from micro structures

This paper concerns orthogonal corner detection from small-scale features (micro-structures) in images, and its use in the registration of an unregistered aerial orthophoto with a LiDAR point cloud data. First, we classify the corners in the aerial and LiDAR images using a corner detection algorithm specifically adapted for orthogonal corner detection. Then we use a search algorithm to register the orthogonal corners from the orthophoto image over the LiDAR image. The search algorithm is tailored for the matchings of orthogonal corners, for which there may be many features in one image with no corresponding feature in the other, but for which there are several corresponding corner features which are highly localized in spatial coordinates. The search algorithm requires an initial coarse estimate of the registration homography in order to perform quickly. We use this method to register a LiDAR and an aerial image of a coastal region with several man-made buildings in the vicinity (and thus well-defined orthocorners). The methods of this experiment can be useful for registering the images from two different modalities, even from high altitude photographs. The experimental results show a high degree of accuracy in the shoreline registration can be attained. The algorithm gives a registration which matches the micro-structures with pixel level accuracy. This is in contrast to the coarse registration, which in our cases was computed from an estimate of the shorelines in the images, which has an error around 10-20 pixels.