Real-time dense 3D reconstruction and camera tracking via embedded planes representation
- PDF / 3,451,651 Bytes
- 12 Pages / 595.276 x 790.866 pts Page_size
- 7 Downloads / 204 Views
ORIGINAL ARTICLE
Real-time dense 3D reconstruction and camera tracking via embedded planes representation Yanping Fu1 · Qingan Yan2 · Jie Liao1 · Alix L. H. Chow3 · Chunxia Xiao1
© Springer-Verlag GmbH Germany, part of Springer Nature 2020
Abstract This paper proposes a novel approach for robust plane matching and real-time RGB-D fusion based on the representation of plane parameter space. In contrast to previous planar-based SLAM algorithms estimating correspondences for each planepair independently, our method instead explores the holistic topology of all relevant planes. We note that by adopting the low-dimensionality parameter space representation, the plane matching can be intuitively reformulated and solved as a point cloud registration problem. Besides estimating the plane correspondences, we contribute an efficient optimization framework, which employs both frame-to-frame and frame-to-model planar consistency constraints. We propose a global plane map to dynamically represent the reconstructed scene and alleviate accumulation errors that exist in camera pose tracking. We validate the proposed algorithm on standard benchmark datasets and additional challenging real-world environments. The experimental results demonstrate its outperformance to current state-of-the-art methods in tracking robustness and reconstruction fidelity. Keywords 3D reconstruction · RGB-D reconstruction · Camera tracking
1 Introduction We are witnessing the vigorous progress of 3D reconstruction [4,6,25,38–40,43,44] in a variety of applications, such as mixed reality, CAD manufacturing, 3D printing and robot perception, especially with the ubiquity of RGB-D sensors Electronic supplementary material The online version of this article (https://doi.org/10.1007/s00371-020-01899-1) contains supplementary material, which is available to authorized users.
B
Chunxia Xiao [email protected] Yanping Fu [email protected] Qingan Yan [email protected] Jie Liao [email protected] Alix L. H. Chow [email protected]
1
School of Computer Science, Wuhan University, Wuhan, Hubei, China
2
Silicon Valley Research Center, JD.com, Mountain View, CA, USA
3
Xiaomi Technology Co. LTD, Beijing, China
and powerful mobile devices. However, accurate 3D shape reconstruction is still quite challenging. It is not easy to design general-purpose reconstruction algorithms to serve various objects and expect all reconstructions to be in high quality. For example, the primitive shapes for offices and trees are disparate. In order to effectively acquire high fidelity geometries, additional domain-specific knowledge is more desirable. Coplanarity serving as a specialized prior has shown compelling capability in RGB-D structural modeling, as flat surfaces are ubiquitous in reality. It provides a high-level constraint for settling the challenge of drift in pose estimation in addition to ICP registration [3]. Although planar-based reconstruction is a promising direction, we have yet to see a simple and holistic solution for the problem of establishing putative plana
Data Loading...
