FFT-Based Terrain Segmentation for Underwater Mapping
Bertrand Douillard · Navid Nourani-Vatani · Matthew Johnson-Roberson · S. Williams · Christopher Roman · Oscar R. Pizarro · Ian Vaughn · Gabrielle Inglis
TL;DR
Segmenting seafloor scans by separating low-frequency ground from objects in the Fourier domain yields object features that support navigation and scan alignment.
BRIEF
Underwater robots build 3D scans of the seafloor, but picking out individual objects from uneven, unstructured terrain is hard. This work turns each scan into an elevation map and analyses it with a fast Fourier transform. Slowly varying ground undulations sit at low frequencies, so the ground is separated by choosing a cut-off frequency, found automatically by peak detection. Points above the estimated ground are clustered into object segments. Tested against hand-labelled data and in alignment experiments, it was compared with three other methods and produced features useful for navigation.
Abstract
A method for segmenting three-dimensional scans of underwater unstructured terrains is presented. Individual terrain scans are represented as an elevation map and analysed using fast Fourier transform (FFT). The segmentation of the ground surface is performed in the frequency domain. The lower frequency components represent the slower varying undulations of the underlying ground whose segmentation is similar to de-noising / low pass filtering. The cut-off frequency, below which ground frequency components are selected, is automatically determined using peak detection. The user can specify a maximum admissible size of objects (relative to the extent of the scan) to drive the automatic detection of the cut-off frequency. The points above the estimated ground surface are clustered via standard proximity clustering to form object segments. The approach is evaluated using ground truth hand labelled data. It is also evaluated for registration error when the segments are fed as features to an alignment algorithm. In both sets of experiments, the approach is compared to three other segmentation techniques. The results show that the approach is applicable to a range of different terrains and is able to generate features useful for navigation.