A geometrical approach to the surface-fitting problem in underwater three-dimensional acoustic images

This paper presents a technique for the segmentation of three-dimensional images acquired by an acoustic camera. The proposed algorithm aims at identifying reliable data derived from objects present in an underwater scene in order to determine elementary components. In this way, the estimation of geometrical parameters for each component is performed with the aim of subsequently performing object-recognition and reconstruction phases. The algorithm is composed of a series of processing steps, starting from a coarse segmentation stage, wherein the most reliable image points likely to correspond to man-made objects are identified, and leading to a refinement phase, in which accurate re-segmentation is performed, to accurately determine points belonging to the same surface. Once objects' components have been identified, geometrical parameters are estimated in order to build an adequate data structure that is easily comparable with a model database. In other words, it exploits the geometrical properties embedded in the sparse and noisy three-dimensional information to group the points better by fitting the current quadric surface. This algorithm can be applied for the reconstruction of virtual environments from acoustic data aimed at robotic applications (e.g. vehicle navigation) and was actually used in an off-shore application consisting of the recognition of the scene seen by an underwater vehicle navigating close to an oil rig. Results on synthetic images proved the goodness and the accuracy of the approach and real examples are also provided to prove the robustness and accuracy of the method.