Visual localization is a critical capability for autonomous systems, enabling them to accuratelyestimate their position and orientation within an environment using visual data. This thesis focuses on achieving a robust and reliable visual localization on both local and global level to enhance localization performance in a wide range of environments. For global localization, we consider the task of both Visual Place Recognition (VPR) and cross-view geo-localization (CVGL). Visual place recognition enables the system to recognize previously visited locations and refine its position estimates. Additionally, it can provide a prelim- inary position estimate when the GPS signal is weak and cannot offer real-time GPS positioning. In this thesis, we leverage high-level semantic information generated from scene-graph generators with traditional VPR pipeline to achieve a more robust global visual localization under extremely diverse and challenging environmental conditions. Cross-view geo-localization is an essential aspect of visual localization, allowing systems to estimate their position and orientation by matching ground-level images to aerial or satellite im- agery. CVGL remains extremely challenging due to the drastic appearance differences across aerial–ground views. In existing methods, the interactive benefits global representations of dif- ferent views are seldom taken into account. In this thesis, we present a novel approach using cross-view knowledge generative techniques in combination with transformers, namely mutual generative transformer learning (MGTL), for CVGL. For local localization, we focus on Visual odometry (VO). Visual odometry is a key compo- nent in visual localization, as it estimates the relative motion of the autonomous system using consecutive visual frames. Existing VO techniques are prone to accumulated error stemming from steering angle deviation, resulting in sub-optimal precision. In this thesis, we propose a novel VO framework comprising steering angle-weighted learning and triple-frame hybrid constraint learn- ing, alleviating the aforementioned problem and achieving a more robust local localization.

Creator

• Zhao, Pengbo

DOI

• https://doi.org/10.21985/n2-w2dh-m526

Subject

• Electrical engineering

Language

• en

Alternate Identifier

• http://dissertations.umi.com/northwestern:16532
• etdadmin_upload_984400

Keyword

• Visual Odometry
• Visual Localization
• Visual Place Recognition
• Cross-View Geo-Localization

Date created

• 2023-01-01

Resource type

• Dissertation

Rights statement

• In Copyright