A new tracking technique for sonar images relying on symbolic information matching through genetic algorithms
Résumé
Sea-bed sonar images of a same 'featureful' area generally need to be registrated in order to be compared because of the different acquisition viewpoints which relative configuration is not a priori known. Acheiving such a registration process may be crucial to perform local comparisons in order to extract, for instance, complementary information about an object observed various times (useful for 3D object reconstruction). Moreover, wider data fusion schemes also require such matching processes to allow the construction and update of representations (databases or maps) in which are merged multiple acquisitions of the sea-bottoms to be described. Positionning systems return efficient global location information but experimental measures show the remaining error needs to be reduced to build reliable data fusion mechanisms. Thus, in this paper, we take advantage of overlapping sea-bed images to provide reference points for high precision relative positionning. Pixel-based registration of sonar images is very time and memory consuming. Moreover, it strongly depends on sonar characteristics. Consequently, our approach consists of a two-step registrating system : its first stage is acquisition-dependent and extracts information through processes tailored for each sonar type. Then, this symbolic knowledge (contours, relevant contacts) feeds a generic system whose goal is to perform the matching task between current information and previously recorded data (coming from the same track in case of overlapping view sonar system, or from databases for older acquisitions). In this study, we focus on the first kind of situation with local data matching in order to estimate relative vehicule motion using consecutive and overlapping images of the covered area. To do so, we implement a matching mechanism using genetic algorithms with variable-length chromosomes coding a field of transformations which size is not a priori known. This extra-flexibility allows to registrate various acquisitions covering a same sea-bed region with multiple distortions, previous algorithms only dealt with single distortion mechanism. When working on non-uniform sea-bed areas, this ground following approach provides additional support in relation to standard navigation system.