Terrain-relative navigation for autonomous underwater vehicles

For map-matching navigation of autonomous underwater vehicles, the central problem is localizing a sonar (or video) image of the seafloor in an existing global bathymetric map. The terrain-relative navigation methodology presented is grounded in the Kalman filter framework and uses its three basic steps: prediction, measurement, and update. In the prediction step, the vehicle location with uncertainty (state and state covariance) is estimated by using the previous vehicle location and dead reckoning. In the measurement step, the current sonar image is matched to the map using the mean absolute difference dissimilarity parameter. Because of the potential for large uncertainties in both the depth values of the image and those of the map, a suite of good matches is accepted rather than the single best match. The location of each match is considered a measurement of the vehicle position. Finally, each match (measurement) is weighted probabilistically using data association techniques, and the resulting best estimate and associated uncertainties comprise the navigation update. This updated position is then used as a basis for the next prediction step, and the process is repeated. Using real bathymetric data from a section of seafloor near the Mid-Atlantic Ridge between the Kane and Atlantis transforms, a simulation showing the success of this algorithm is presented.