A Processing Pipeline for Descriptive Underwater 3D Occupancy Mapping with Scanning Sonar

We discuss an approach for 3D mapping using a remotelyoperated vehicle (ROV) in cluttered shallow-water environments where port and harbor infrastructure inspection is of interest. Our goal is to equip the ROV with a mapping method that will support sound decision-making in the process of autonomously exploring a a priori unknown environment with a scanning sonar. This capability will be important in the initial phases of inspecting an unstructured environment, in which obtaining situational awareness, to an extent that permits reasoning about obstacles and collision avoidance, is an important first step prior to the detailed inspection of specific areas. One of the most challenging aspects of this task is producing a descriptive map amidst the high levels of noise present in the sonar data. High-accuracy underwater sonar-based mapping has been achieved using variants of the iterative closest point (ICP) scan-matching algorithm [8], which have been applied in port and harbor settings to produce comprehensive 3D point clouds [1], [5]. In addition, particle filters have been applied [7] to produce 3D occupancy grid maps in settings with substantial supporting structure, such as underwater caves [3] and cisterns [10]. For 3D reasoning about maneuvering and obstacle avoidance, an accurate grid-based map is highly valuable. It is our goal to produce such maps in the unstructured environments in which scan-matching has previously succeeded in producing high-quality point cloud maps. To do so, we will leverage recent work in 3D occupancy mapping [9] that allows predictive inference to be performed over sparse and noisy data, producing a map that will serve