Bundle adjustment from sonar images and SLAM application for seafloor mapping

Abstract

This paper reports on two-view bundle adjustment using sonar images, specifically focusing on feature detection and a sensor measurement model for imaging sonar. To overcome limited sensor information for underwater navigation, we use Dual frequency IDentification SONar (DIDSON) in the imaging mode to provide spatial constraints when a scene is revisited. Unlike terrestrial images, sonar images are usually low resolution with highly speckled noise. We found that exploiting features from nonlinear scale space improves feature detection. In this paper, we adopt KAZE features and use random sample consensus (RANSAC) to refine correspondences. Using these correspondences, we propose point-based relative pose estimation via bundle adjustment. The target application that this work focuses on is underwater seafloor mapping, and the proposed model assumes a fixed elevation. Through this work, we present (i) validation of nonlinear scale space features for sonar images and (ii) proposal of a sonar sensor measurement model for underwater simultaneous localization and mapping (SLAM). The proposed method will be validated through both synthetic data sets and a tank test for seafloor mapping.