Scalable Visual Localization with Dense Geometry Modelling For Autonomous Navigation

The Hong Kong University of Science and Technology
Department of Computer Science and Engineering


PhD Thesis Defence


Title: "Scalable Visual Localization with Dense Geometry Modelling For 
Autonomous Navigation"

By

Mr. Huaiyang HUANG


Abstract

State estimation is a crucial building block for an autonomous navigation 
system with high intelligence. Techniques for state estimation enable an 
agent to perceive geometric information related to itself and its 
surroundings, e.g., sensor pose tracking, self-localization, and mapping. 
For run-time robustness and precision, generally, robots are equipped with 
heterogeneous sensors to guarantee necessary information redundancy, which 
requires the state estimation module to fully exploit the advantages of 
individual sensors while fusing the perceptual data in a complementary 
manner.

This thesis presents a scalable state estimation system via dense geometry 
modelling and cross-modal visual localization We begin with reconstruction 
from ranging sensors that provide the dense geometric structure for the 
localization system. We first propose a multiview point cloud registration 
method with point-set bundle adjustment, which simultaneously refines the 
poses of individual scans in an alternating optimization framework. With a 
deformable map representation, we further propose an efficient and 
globally consistent odometry and mapping system for typical LiDARs, which 
is capable of providing geometric priors in environments with different 
scales.

In the next stage, we explore the cross-modal localization method as an 
intermediate way of combining the advantages of visual and geometric 
measurements. The basic framework associates the sparse visual structure 
with the pre-built geometric structure, and then introduces the structure 
regulation in the optimization to eliminate the drift and align the 
coordinate globally. To this end, we propose a monocular localization 
system based on the Signed Distance Field (SDF), which could robustly 
initialize with the pre-built dense map and consistently track the camera. 
Finally, to make the system more applicable, we further model the scene 
geometry as Gaussian Mixture Model (GMM) and introduce the geometric 
information into the visual localization system. Overall, this thesis 
presents a scalable pipeline that reconstructs the geometry of the 
environment and then tracks cameras over the pre-built dense map.


Date:			Tuesday, 29 November 2022

Time:			3:00pm - 5:00pm

Venue:			Room 4472
 			lifts 25/26

Chairperson:		Prof. Sen YANG (PHYS)

Committee Members:	Prof. Ming LIU (Supervisor)
 			Prof. Qifeng CHEN
 			Prof. Pedro SANDER
 			Prof. Lujia WANG (ECE)
 			Prof. Hesheng WANG (Shanghai Jiao Tong University)


**** ALL are Welcome ****