Pose estimation and trajectory derivation from underwater imagery

Obtaining underwater imagery is normally a costly affair since expensive equipment such as multi-beam sonar scanners need to be utilized. Even though such scanners provide imagery in form of 3D point clouds, the tasks of locating accurate and dependable correspondences between point clouds and registration can be quite slow. Registered 3D point clouds can provide pose estimation and trajectory information vital to the navigation of a robot, however, the slow speed of point cloud registration normally means that maps are generated offline for later use. Furthermore, any algorithm must be robust against artifacts in 3D range data as sensor motion, reflection and refraction are commonplace. In our work we describe the use of the SIFT feature detector on scaled images based on point clouds captured by sonar in order to register them in real-time. This online registration approach is used to derive navigational information vital to underwater vehicles. The algorithm utilizes the known point correspondence registration algorithm in order to achieve real-time registration of point clouds, thereby generating 3D maps in real-time and providing 3D pose estimation and trajectory information.