Robust Global Structure From Motion Pipeline With Parallax on Manifold Bundle Adjustment and Initialization
| Autor: | Brenton Leighton, Gamini Dissanayake, Liyang Liu, Liang Zhao, Teng Zhang, Shoudong Huang |
|---|---|
| Rok vydání: | 2019 |
| Předmět: |
Control and Optimization
Computer science business.industry Mechanical Engineering ComputingMethodologies_IMAGEPROCESSINGANDCOMPUTERVISION Biomedical Engineering Initialization Bundle adjustment Simultaneous localization and mapping Computer Science Applications Human-Computer Interaction Artificial Intelligence Control and Systems Engineering Robustness (computer science) Computer vision Computer Vision and Pattern Recognition Artificial intelligence business Parallax ComputingMethodologies_COMPUTERGRAPHICS |
| Zdroj: | IEEE Robotics and Automation Letters. 4:2164-2171 |
| ISSN: | 2377-3774 |
| DOI: | 10.1109/lra.2019.2900756 |
| Popis: | © 2016 IEEE. In this letter, we present a novel global structure from motion (SfM) pipeline that is particularly effective in dealing with low-parallax scenes and camera motion collinear with the features that represent the environment structure. It is therefore particularly suitable in Urban SLAM, in which frequent road-facing motion poses many challenges to conventional SLAM algorithms. Our pipeline includes a recently explored bundle adjustment (BA) method that exploits a feature parameterization using Parallax angle between on-Manifold observation rays (PMBA). It is demonstrated that this BA stage has a consistently stable optimization configuration for features with any parallax and therefore low-parallax features can stay in reconstruction without pre-filtering. To allow practical usage of PMBA, we provide a compatible initialization stage in the SfM to initialize all camera poses simultaneously, exhibiting friendliness to collinear motion. This is achieved by simplifying PMBA into a hybrid graph problem of high connectivity yet small node set size, solved using a robust linear programming technique. Using simulations and a series of publicly available real datasets including "KITTI" and "Bundle Adjustment in the Large," we demonstrate the robustness of the position initialization stage in handling collinear motion and outlier matches, superior convergence performance of the BA stage in the presence of low-parallax features, and effectiveness of our pipeline to handle many sequential or out-of-order urban scenes. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|