Title:
Improving the applicability of visual SLAM with submodular submatrix selection

Thumbnail Image
Author(s)
Zhao, Yipu
Authors
Advisor(s)
Vela, Patricio A.
Advisor(s)
Editor(s)
Associated Organization(s)
Series
Supplementary to
Abstract
The objective of the thesis is to improve the applicability of Visual Simultaneous Localization and Mapping (VSLAM) on diverse platforms and scenarios, which has broad impact on practical applications in Robotics and Argumented Reality (AR). Traditionally, a large fraction of effort on VSLAM has focused on performance, for instance accurate pose tracking, dense mapping, etc. The computation cost of VSLAM, on the other hand, is commonly overlooked: many VSLAM systems have to run on desktop CPUs or even GPUs to meet real-time requirements. Until very recently, the applicability of VSLAM draws the attention of community, with target applications on diverse platforms (e.g. micro flying vehicles, AR headset) and scenarios (e.g. low-texture, fast motion). However, state-of-the-art applicable VSLAM involves design choices that trade efficiency with significant sacrifice of performance, therefore with low tracking accuracy and high sensitivity to working environment. In this thesis, we study feature-based BA SLAM, which has high performance in general but also high computation cost. A series of improvements are proposed to improve both the efficiency and performance of feature-based BA SLAM for diverse platforms and scenarios. As recognized in the SfM and SLAM community, the structure of the SLAM problem can be represented with two equivalent representations: factor graph and Jacobian matrix. From the perspective of information preservation, the full factor graph that contains many inter-connections between nodes should be used. However, for a real-time applicable SLAM, a small and sparse graph (Jacobian) is preferred. A rich body of work has explored offline or posterior graph sparsification. Instead, the scope of this thesis is on online graph selection and sparsification. The thesis is based upon theorems developed in the submatrix selection literature. Originating from computational theory and machine learning, submatrix selection aims at identifying a subset of columns/rows from the original matrix, while maximizing matrix revealing metrics such as the Frobenius Norm and logDet. An optimally selected submatrix not only preserves the most information from original matrix but is also much smaller and sparser than the original one. Small-size and sparsity are preferred for efficient numerical optimization; the performance-efficiency trade-off of optimization-related process such as VSLAM is improved thereafter. In Chapter 3, submatrix selection is introduced to guide the feature matching effort in the feature matching module of the VSLAM front-end. Chapter 4 extends the concept of submatrix selection to submatrix tuning, for improving the conditioning of the line-assisted VSLAM. In Chapter 5, the local map data structure and data selection process are explored, as it plays a critical role in VSLAM front-end robustness. Submatrix selection enables efficient construction and querying of a compact local map. In Chapter 6, submatrix selection is introduced to the BA-based VSLAM back-end, which results in a cost-effective back-end solution with superior performance when running on compute limited devices. Finally Chapter 7 explores VSLAM in mobile robotic systems for closed-loop and online usage. VSLAM pose estimation is integrated with high-rate inertial data to generate feedback control signal, which is crucial for close-loop navigation. The benefit of low-latency VSLAM is revealed in the closed-loop navigation, which is a major application of VSLAM.
Sponsor
Date Issued
2019-08-21
Extent
Resource Type
Text
Resource Subtype
Dissertation
Rights Statement
Rights URI