OMNIDIRECTIONAL PERCEPTION: ADVANCEMENTS IN ROBUST VISUAL LOCALIZATION AND MAPPING

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


PhD Thesis Defence


Title: "OMNIDIRECTIONAL PERCEPTION: ADVANCEMENTS IN ROBUST VISUAL LOCALIZATION
AND MAPPING"

By

Mr. Huajian HUANG


Abstract

Effective scene perception forms the bedrock of robots and intelligent systems,
enabling them to construct accurate maps of their surroundings and localize
within them. Despite substantial progress, achieving robust localization and
generating suitable environment representations for high-level tasks remains a
challenge.

In this thesis, we address these challenges through a series of innovative
contributions. To counter the fragility of monocular SLAM systems, we introduce
a recovery mechanism that treats pose estimator failures as stochastic
processes. This mechanism allows for rapid system reinitialization and map
reintegration, enhancing robustness in navigation scenarios without revisiting
locations.

Recognizing the limitations of traditional cameras' restricted fields of view,
we harness the capabilities of 360-degree cameras to enable comprehensive
perception. Our novel approach, 360VO, employs a single 360-degree camera for
direct visual odometry, optimizing photometric residuals to achieve reliable
localization and semi-dense point cloud mapping.

Expanding beyond geometric mapping, we propose a method to enhance path
planning and navigation by abstracting topological graphs that capture scene
relationships. It is achieved by exploiting stable landmark co-visibility in
omnidirectional images and estimating semantic coefficients to discern
topological relationships. This approach seamlessly integrates into
omnidirectional visual SLAM, offering computational efficiency.

In the pursuit of real-time immersive exploration in indoor environments, we
delve into omnidirectional radiance fields. By exploiting the capacity of
positional encoding and neural networks in a geometry-aware fashion, we
heighten rendering speed and recover high-frequency details. With floorplan
guidance, our system is capable of delivering an appealing and immersive indoor
roaming experience.

Furthermore, since robotic systems generally operate in diverse and dynamic
environments, detecting and discarding dynamic elements could enhance system
robustness. Therefore, we introduce an efficient long-term visual tracker that
leverages cross-level feature correlation and adaptive tracking. Importantly,
we proposed a novel omnidirectional tracking benchmark dataset, referred to as
360VOT. The assessment of prevailing tracking algorithms originally designed
for perspective tracking underscores the challenges and opportunities within
omnidirectional object tracking.

In conclusion, this thesis advances visual localization and mapping
methodologies, propelling the field toward omnidirectional perception. Our
innovative approaches, centered around omnidirectional images, contribute to
enhanced system performance while acknowledging emerging challenges. This
pursuit of omnidirectional perception holds significant promise within the
realms of computer vision and robotics.


Date:                   Monday, 4 December 2023

Time:                   2:30pm - 4:30pm

Venue:                  Room 4502
                        Lifts 25/26

Chairman:               Prof. Jianan QU (ECE)

Committee Members:      Prof. Sai Kit YEUNG (Supervisor)
                        Prof. Chi Keung TANG
                        Prof. Dan XU
                        Prof. Rob SCHARFF (ISD)
                        Prof. Jia PAN (HKU)


**** ALL are Welcome ****