Within the community of laser-based range sensing, specialized algorithms have been
designed to generate watertight, 3D mesh models from high-resolution point clouds [74], [45]. Laser-based range sensing, ubiquitous in ground, air, and space applica-
tions, however, yields substantially higher-resolution point clouds than does underwa- ter acoustic range sensing: typically sub-millimeter versus sub-decimeter resolution.
This is evident in several studies that have pursued mapping of 3D structures using underwater acoustic range data [29], [57], [94], [25]. Fortunately, a number of deriva-
(a) Photo and point cloud from the USNS Red Cloud. The photo shows the Red Cloud at right, with a ship of the same class departing at left. The point cloud shows a rudder and portions of both propellers.
(b) Photo and point clouds from the USCGC Venturous. The point clouds show the Venturous from the starboard side and the stern, respectively. Photo credit: US Coast Guard, http://www.uscg.mil/lantarea/cgcventurous/
(c) Photo and mesh from the SS Curtiss. This mesh is based on high-quality, com- prehensive point cloud data and was used as one of the primary tools in planning algorithm development.
Figure 6-1: A summary of HAUV field experiments performed in support of coverage algorithm development and planned path execution, part one.
(a) Photo and mesh from the Nantucket Lightship. This small ship was used for practicing the execution of a planned inspection route.
(b) Photo and mesh from the M/V Terry Bordelon. The mesh depicted focuses on a propeller and its supporting structures. This small ship was used for testing the production of an improved-resolution mesh after executing a planned inspection. Photo Credit: Bordelon Marine, http://www.bordelonmarine.com/terry.html
(c) Photo and mesh from the USCGC Seneca. This is the only vessel that was visited for a second field test. The mesh, developed from the first test, was used to plan a coverage path that was executed during the second test.
Figure 6-2: A summary of HAUV field experiments performed in support of coverage algorithm development and planned path execution, part two.
outliers [166], [78], and these provide a direct avenue for us to pursue our identification
survey mesh model.
Figures 6-3 and 6-4 illustrate the execution and processing of an identification
survey from start to finish. First, the HAUV traces out the walls of a safe bounding box that observes the stern from a distance known to be collision-free; thousands
of DIDSON frames are collected along with navigation estimates. Evident in the sonar frames shown is the range noise which makes this modeling task difficult in
comparison to laser-based modeling.
To transform a set of dense, raw-data point cloud slices into a 3D mesh recon-
struction, we first apply a simple outlier filter to the individual sonar frames collected. All points of intensity greater than a specified threshold are introduced into a slice,
and then each is referenced using the HAUV’s seafloor-relative navigation. These steps are performed using software from SeeByte Ltd., and all remaining steps are
performed using Meshlab [41]. These and other software tools used for processing and acquisition of data are described in Table B.3 in Appendix B. After assembling the
sonar frames into a single point cloud, areas containing obvious noise and second re- turns are cropped out manually. The raw points are then sub-sampled using Poisson
disk sampling [42], which draws random samples from the point cloud, separated by a specified minimum distance. The point cloud is typically reduced to about 10% of
its original density, and it is then partitioned into separate component point clouds.
Partitions are selected based on the likelihood that they will yield individually
well-formed surface reconstructions. Objects such as rudders, shafts, and propellers are thin structures that may not be captured in the final model without separate pro-
cessing from the hull. Normal vectors are computed over the component point clouds, and some flat surfaces, for which only one of two sides was captured in the data, are
duplicated. A point’s normal vector is computed by applying principal component analysis to the point’s k nearest neighbors, and the normal’s direction is selected to
locally maximize the consistency of vector orientation [74]. Both sub-sampling and estimation of normals are key steps in the processing sequence, found in practice to
(a) Survey in progress at the SS Curtiss, with a diagram of the identification survey procedure.
(b) Representative sonar frames from survey of SS Curtiss running gear, looking up at the shaft and propeller. Ranges are given in meters.
Figure 6-3: An overview of the identification survey procedure and the data obtained from it, part one.
(a) Raw-data point clouds obtained from the starboard-side wall and bottom wall of the iden- tification survey, respectively.
(b) Merged, subsampled data is displayed with a vertex normal pointing outward from each individual point.
(c) A mesh model of SS Curtiss generated by applying the Poisson reconstruction algorithm to the point cloud of (b).
Figure 6-4: An overview of the identification survey procedure and the data obtained from it, part two.
density, evenly-distributed set of points, and normals aid in defining the curvature of
the surface.
The Poisson surface reconstruction algorithm [91] is next applied to the oriented
point clouds. Octree depth is selected to capture the detail of the ship structures without including excess roughness or curvature due to noise in the data. The com-
ponent surfaces are merged back together, and a final Poisson surface reconstruction is computed over the components. If the mesh is used as a basis for high-resolution
inspection planning, then it may be further subdivided to ensure the triangulation suits the granularity of the inspection task. We iteratively apply the Loop subdivision
algorithm [112] for this purpose, which divides each triangle larger than a specified size into four subtriangles.