Modern visual SLAM (vSLAM) algorithms take advantage of computer vision developments in image processing and in interest point detectors to create maps and trajectories from camera images. Different feature detectors and extractors have been evaluated for this purpose in air and ground environments, but not extensively for underwater scenarios. In this paper (I) we characterize underwater images where light and suspended particles alter considerably the images captured, (II) evaluate the performance of common interest points detectors and descriptors in a variety of underwater scenes and conditions towards vSLAM in terms of the number of features matched in subsequent video frames, the precision of the descriptors and the processing time. This research justifies the usage of feature detectors in vSLAM for underwater scenarios and present its challenges and limitations.
A comparative evaluation of interest point detectors and local descriptors for visual SLAM
