Unconstrained Synthetic Aperture Sonar

点击翻译按钮获取中文摘要

Sonar is an essential sensor for underwater applications as it provides data under bad or even no visibility conditions and over longer distances. But its spatial resolution depends on a combination of transducers to (roughly) approximate sampling beams by interferences. A larger number of transducers placed on a larger area accordingly provide a higher resolution. But the number of transducers in a sonar sensor is limited by many factors like sensor size, power consumption and costs. A popular approach is hence the use of a synthetic aperture, i.e., the sonar with its N transducers is positioned at k places to generate a virtual sensor with kN transducers.The state of the art for this synthetic aperture sonar (SAS) is strongly coupled to constraints on the way it can be used. For example, the k poses often have to be equidistantly placed on a line perpendicular to the sensor. This is motivated by the intention to ease the signal processing as well as by practical aspects: a vehicle, e.g., a surface vessel or an Autonomous Underwater Vehicle (AUV), with a sonar facing down to the sea-floor is only required to navigate with constant speed on a straight line. But it also significantly limits the scope of the vehicle.This project develops the foundations for an unconstrained SAS, i.e., a SAS that a) can be computed on arbitrary trajectories b) without the requirement of navigation sensor data (GPS, INS, etc.). As shown in preliminary work, a registration of the raw scans provides sufficiently precise location estimates to compute a higher-resolution spatial reconstruction of the scene. With respect to a) contributions to the formulation and solution to the unconstrained SAS problem are derived. These are also of interest to related areas, e.g., the use of radar on mobile systems like robots or automobiles, remote sensing with radar, or medical imaging with ultra-sound. With respect to b) methods for robust 2.5D and 3D registration are investigated, especially spectral methods, which are very well suited for raw data that is strongly affected by noise.