Segmentation of bathymetric profiles and terrain matching for underwater vehicle navigation

Abstract

A terrain-based underwater navigation using sonar bathymetric profiles is presented. It deals with matching high-resolution local depth maps against a large on-board reference map. The matching algorithm locates the local depth map within the a priori known larger map to determine the absolute position and heading of the vehicle. Two separate approaches for this problem are presented. The first uses a contour-based representation of depth maps. Contours are extracted from both local and reference maps. Invariant attributes under rigid plane transformation are associated with each contour point, so that the problem is reduced to a point-based matching algorithm: given two point sets, find correspondences and estimate transformation between the two sets. We shall particularly focus on the formalism of partial differential equations, which is used to smooth depth maps in a morphologically invariant way and to obtain anisotropic contours. The second approach is also based on a correspondence algorithm. Here, discriminant parameters are associated with each local map and the correspondence is established in the parameter space. The extraction of discriminant parameters is based on a wavelet packet analysis. Their discriminant power and robustness are analysed. In order to reduce the space of possible robot locations, an area of localization is first delimited by a crude method. A finer interpolation follows, using a neural network. Both methods have been designed keeping in mind limited computational load in order to board future systems on small operational robots and they have been tested using a real terrain model of the Var underwater canyon (France); a multibeam sounder simulator is used in the first approach in order to simulate sonar performances as well as bathymetric image deformations due to errors in short-term navigation.