Polarization-based underwater geolocalization with neural network models

Understanding the sea has always been a core theme of mankind's scientific endeavors. In recent years, the spatial and temporal scope of such exploration has been further expanded thanks to the rapid development of robot technologies. However, due to the unavailability of GPS signals underwater, efficient, high-accuracy and real-time navigation of man-made autonomous devices below the water surface remains an open challenge. Inspired by the visual system of marine animals such as mantis shrimps, polarization-based underwater geolocalization (PUG) aims to solve this problem by leveraging the rich polarization patterns generated by sunlight and moonlight as they are refracted and scattered underwater. Those patterns encode information about the celestial body's location in the sky, and once such information is extracted, one can subsequently combine it with the coordinated date and time when the observation is made to infer the observer's geolocation. In this thesis, we present a deep learning-based approach to PUG, which features higher accuracy, lower latency and better applicability to real-world scenarios than the existing parametric method. As a preliminary step, we develop a calibration algorithm to resolve the inherent angle of polarization (AoP) shift when the polarization camera is equipped with a wide-angle lens. This calibration algorithm enables us to use omnidirectional polarization images for training and evaluating our PUG methods. Next, we establish that rotation invariance and temporal modeling are the keys to designing network models that estimate sun location from omnidirectional polarization images. To this end, we propose RI-ResNet and RDM to address the two key points respectively and construct our learning-based PUG method by joining them with a particle filter (PF) geolocation estimator. We demonstrate that our method can learn to geolocalize in different sites, seasons, water types and depths. We have also proved the feasibility of PUG at nighttime for the first time ever. Finally, we focus our attention on improving model generalizability by proposing SecTran-M, a rotation invariant Transformer network, as the sun localization backbone and using unscented Kalman filter (UKF) for temporal modeling. Evaluation results on cross-site tasks show that our locally-trained model can achieve coarse-level geolocalization on a global scale, taking us one step closer to a globally functional high-accuracy PUG method that we have always envisioned.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Computer Science

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2023

Author dc:creator

• Bai, Xiaoyang

Contributors dc:contributor

• Schwing, Alexander
• Gruev, Viktor
• Forsyth, David
• Laze, Svetlana
• Schechner, Yoav

• Underwater Geolocalization
• Polarization-based Underwater Geolocalization
• Machine Learning
• Computer Vision
• Neural Network

Statement dc:rights

• Copyright 2023 Xiaoyang Bai

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/121977

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F121977