Options
2014
Conference Paper
Titel
Bayesian reconstruction of seafloor shape from side-scan sonar measurements using a Markov Random Field
Abstract
To explore the seafloor, a side-scan sonar emits a directed acoustic signal and then records the returning (reflected) signal intensity as a function of time. The inversion of that process is not unique: multiple shapes may lead to identical measured responses. In this work, we suggest a Bayesian approach to reconstructing the 3D shape of the seafloor from multiple sonar measurements, inspired by the state-of-the-art methods of inverse raytracing that originated in computer vision. The space near the bottom is modelled as a grid of voxels, whose occupancies are represented by random binary variables. Any assignment of occupancies corresponds to some seafloor shape. A global multi-component energy potential describes how well the resulting surface agrees with the sonar data and with the a priori assumptions. Minimization of energy is equivalent to finding the maximum a posteriori (MAP) assignment to this Markov random field (MRF) and is done using the iterated belief propagation (BP) algorithm. The critical step in this method is to compute messages from ""factors"" representing the sonar beams to voxels. Naïvely, its complexity scales exponentially with the number of voxels traversed by a beam. Unlike inverse raytracing, where a pixel value constrains voxels only along a single view ray, a sonar beam involves voxels within a relatively wide cone. Employing dynamic programming techniques and space-filling curves, we were able to develop a practical approximate solution to this problem. The algorithm is not restricted to side-scan sonar reconstruction and could be applied to medical ultrasound or ultra wide-band (UWB) radar imaging.