Active Simultaneous Localization and Mapping in Perceptually Aliased Underwater Environments

Abstract

The problem of semantic simultaneous localization and mapping (SLAM) is especially difficult in underwater environments due to sensor characteristics and terrain. The primary underwater sensor, sonar, is subject to multipath reflections, as well as an elevation angle ambiguity that makes it difficult to integrate its data into SLAM frameworks. Furthermore, the lack of training data makes it difficult to accurately obtain object detections from sonar for semantic, or object-based, SLAM. Finally, the technique of actively choosing trajectories that can take into account data association ambiguities between semantic landmarks is still an open research area. This work comprises of two main contributions: the design and implementation of a Gaussian mixture representation for data association of semantic object detections in environments perceived with sonar, and the design and implementation of a path planning algorithm that allows a vehicle to actively seek trajectories that disambiguate and elucidate the robot's position and its map of the surrounding environment. These two techniques are tested in various experimental settings, with results showing the novel ability to actively navigate with an awareness of semantic object landmarks' data association ambiguities. Future work will involve further experimental evaluation on combining the underwater mapping techniques with the active navigation techniques developed in this thesis, as well as the development of more techniques for designing and training object detectors for sonar.