Abstract
We present a novel, effective method for global point cloud registration problems by geometric topology. Based on many point cloud pairwise registration methods (e.g., ICP), we focus on the problem of accumulated error for the composition of transformations along any loops. The major technical contribution of this paper is a linear method for the elimination of errors, using only solving a Poisson equation. We demonstrate the consistency of our method from the Hodge-Helmholtz decomposition theorem and experiments on multiple RGBD datasets of real-world scenes. The experimental results also demonstrate that our global registration method runs quickly and provides accurate reconstructions.
Introduction
In 3D scene reconstruction from multiple point clouds—such as those captured by handheld scanners—the primary challenge is the accumulation of error, commonly referred to as "drift" or the "loop closure problem." When registering a sequence of clouds (A to B, B to C, and C back to A), the final transformation rarely aligns perfectly with the starting point. Traditional sequential methods exacerbate this drift, while existing global methods often rely on computationally expensive non-linear optimizations or rough approximations like rotation averaging. This paper introduces a topological approach to achieve global consistency by linearizing the problem through Lie-algebraic cohomology.
Methodology
The proposed method shifts the focus from iterative optimization to topological decomposition. The pipeline involves four key stages:
1. From Graphs to Complexes
Rather than treating the data as a simple graph of nodes and edges, we represent the collection of point clouds as a Cech complex. This allows us to capture higher-order topological information, such as the mutual intersection of three or more point clouds, represented as 2-simplices or "faces."
2. Linearizing Motion in Lie Algebra
Rigid motions (rotations and translations) are inherently non-linear. By mapping these transformations into their corresponding Lie algebra ($so(3) \times \mathbb{R}^3$), they become a "simplicial 1-form." In this linearized space, the motions act as a vector field on the graph, allowing for linear operations.
3. Hodge-Helmholtz Decomposition
The core of the method is the application of the Hodge-Helmholtz decomposition theorem. Any vector field on a complex can be decomposed into three components: an exact form (curl-free), a co-exact form (divergence-free), and a harmonic form. In this context, the exact component represents the "true," globally consistent poses, while the remaining components represent noise and accumulated errors.
4. The Poisson Equation Shortcut
By applying the co-boundary operator to the 1-form, we derive a Poisson equation. Solving this linear system allows for the direct extraction of the exact component, effectively "cleaning" registration errors across the entire viewing graph in a single step. To maintain efficiency, the authors use graph embedding to map the viewing graph onto a 2D surface, creating a manageable 2-skeleton for the decomposition.
Experimental Results
The method was evaluated on the ICL-NUIM (synthetic) and 7-Scenes (real-world) datasets.
- Accuracy: The "Hodge" method consistently outperformed BundleFusion (BF), a current industry standard. In some cases, it produced sharper reconstructions than the provided ground truth by correcting simulated sensor noise.
- Convergence: Loop consistency error ($D$) typically drops to effectively zero within two to three iterations (e.g., from $1.04 \times 10^{-1}$ to $0.00$).
- Robustness: The method successfully reconstructed scenes where BundleFusion failed, such as the "office2" scene in the ICL-NUIM dataset.
Discussion and Limitations
The primary insight of this work is that global consistency is mathematically equivalent to the "exactness" of a 1-form in cohomology. This shifts the computational burden from heavy non-linear solvers to a single linear system.
However, the method relies on a "near identity" assumption: relative transformations between adjacent clouds must be relatively small for the Lie-algebraic linear approximation to remain accurate. Furthermore, it requires a reasonably accurate initial alignment (e.g., from ICP) to construct the initial 1-form. Future work will focus on handling sparse poses with larger transformations and exploring more robust topological structures.
Summary updated