Date of Award

2024

Degree Type

Dissertation

Degree Name

Doctor of Philosophy in Ocean Engineering

Department

Ocean Engineering

First Advisor

Mingxi Zhou

Abstract

The polar oceans are crucial in influencing global overturning circulation, biological carbon pumps, and the carbon cycle. Unmanned Underwater Vehicles (UUVs) are promising tools to explore this area for scientific sampling and monitoring. However, underwater localization is challenging because of the lack of GPS. To improve the robustness of UUV under-ice navigation ability, in this dissertation, we investigated a multi-sensor fusion localization system for ice-water inter-face exploration. Specifically, the state-of-the-art Multi-State Constraint Kalman Filter (MSCKF) is leveraged to perform state estimation using Doppler Velocity Log (DVL), Pressure, Inertial Measurement Unit (IMU), Forward-looking Sonar (FLS) and Mono-camera. Furthermore, robust feature estimation from both the camera and FLS are present to perform localization under challenging under-ice environments (e.g., degenerated image conditions and motions).

In this study, a portable ROV is constructed with a suite of underwater navigation sensors for study the under-ice environments. Field trials have been conducted under freshwater ice (Michigan frozen river) and seawater ice (Utqiagvik, Alaska). For better localization accuracy, the first of its kind for underwater scenarios, hardware time synchronization for multiple devices is demonstrated. A robust Dead-reckoning is presented using MSCKF with system initialization using IMU, DVL and Pressure for the dynamic environment.

At close distances (1-2m), such as surveying the ice-water interface, degraded conditions may occur. But, the camera at a close distance could provide motion cues to improve dead-reckoning navigation based on IMU, DVL and pressure sensors. To this end, this dissertation introduced a tightly-coupled visual-aided odometry method with two new mechanisms to improve the localization robustness. First, a modified keyframe-based feature marginalization for the MSCKF feature update is presented to improve the feature position estimation and reduce the computational cost. Second, the sparse DVL point cloud (e.g., 4 points) and visual feature association and feature enhancement are presented. The proposed technique is validated using an under-ice dataset with significant improvement found compared to the dead-reckoning solution. More importantly, we found that improvements in localization accuracy when the new mechanism compared to the conventional VIO method.

Furthermore, the under-ice localization framework is extended to include FLS which operates effectively in turbid water and has been utilized in various under-water applications. The thesis advances the field FLS-aided odometry field by introducing a new FLS feature position estimation method to resolve the elevation ambiguity problem with feasible assumptions. The new method overcomes the issue of inaccurate feature triangulation caused by severe degenerated motion, such as 2D plane motion. Specifically, the method constructs a submap from a series of point clouds estimated using the leading-edge features from the FLS images. Then, a plane model is created to fit the submap, then elevation angle of tracked features in the FLS images is quickly solved using Trigonometric Identities. Based on the estimated elevation angle and the range and azimuth angle provided by the sonar image, we could compute the 3D position of the feature which will be used for vehicle pose update. The proposed method is evaluated based on the under-ice dataset and improvements were found by comparing it with the visual-DVL point cloud fused method.

DP-VIO.gif (46725 kB)
DP-VIO Video

DPS-VIO.gif (90881 kB)
DPS-VIO Video

Available for download on Friday, September 12, 2025

Share

COinS