Strategies for filtering incorrect matches in seabed mosaicking

Monitoring animal populations in benthic habitats is essential to detecting changes in the local ocean environment and marine ecosystem. Images of animals on the seafloor are obtained by drop-cameras or digital still-cameras mounted on Remotely Operated Vehicles (ROVs). Population statistics have a wide range of applications in the fishery industry, oceanographic research (e.g. population studies, habitat analysis), as well as for the oil and gas industry (e.g. population monitoring for environmental impact assessment). Most ROV imaging transects deliberately produce overlap between successive or adjacent images, such that individual animals could appear in several images, which could yield inaccurate counts. In order to eliminate the possibility of counting the same animal more than once, the overlap between images must be detected and cropped from one of the images. We are developing a feature-based mosaicing algorithm that uses Scale Invariant Feature Transform (SIFT) features in which feature descriptors of images are extracted and appropriate correspondences are found and matched by computing the Standardized Euclidean distance between descriptor vectors. The Homography matrix between each pair of images is then estimated by RANdom SAmple Consensus (RANSAC). Finally, by using the estimated Homography the final mosaic is generated with a multi-band blending algorithm.