Parallax Bundle Adjustment on Manifold with Improved Global Initialization

EasyChair Preprint 748

Abstract

In this paper we present a novel extension to the parallax feature based bundle adjustment (BA). We take parallax BA into a manifold form (PMBA) along with an observation-ray based objective function. This formulation faithfully mimics the projective nature in a camera's image formation, resulting in a stable optimization configuration robust to low-parallax features. Hence it allows use of fast Dogleg optimization algorithm, instead of the usual Levenberg Marquardt. This is particularly useful in urban SLAM in which diverse outdoor environments and collinear motion modes are prevalent. Capitalizing on these properties, we propose a global initialization scheme in which PMBA is simplified into a pose-graph problem. We show that near-optimal solution can be achieved under low-noise conditions. With simulation and a series of challenging publicly available real datasets, we demonstrate PMBA's superior convergence performance in comparison to other BA methods. We also demonstrate, with the "Bundle Adjustment in the Large" datasets, that our global initialization process successfully bootstrap the full BA in mapping many sequential or out-of-order urban scenes.

Keyphrases: Global SfM, Mapping, Monocular SLAM, bundle adjustment, computer vision, global initialization, parallax bundle adjustment

@booklet{EasyChair:748,
  author    = {Liyang Liu and Teng Zhang and Yi Liu and Brenton Leighton and Liang Zhao and Shoudong Huang and Gamini Dissanayake},
  title     = {Parallax Bundle Adjustment on Manifold with Improved Global Initialization},
  doi       = {10.29007/jvg7},
  howpublished = {EasyChair Preprint 748},
  year      = {EasyChair, 2019}}