Large Common Plansets-4-Points Congruent Sets for Point Cloud Registration

Cedrique Fotsing, Nafissetou Nziengam, Christophe Bobda

TL;DR

This paper introduces LCP-4PCS, a novel point cloud registration method that leverages plane segmentation to improve alignment accuracy and efficiency, especially for datasets with low overlap. By focusing on aligning the largest common planes rather than just points, LCP-4PCS significantly reduces computation time and yields better results than traditional 4PCS.

What Problem Does This Paper Address?

Accurately merging multiple 3D scans (point clouds) into a single, consistent model is crucial for applications like civil engineering, cultural heritage preservation, and robotics. This process, known as point cloud registration, involves aligning scans that are often captured from different viewpoints and may not have significant overlap. Existing methods, like the widely used 4PCS (4-Points Congruent Sets), struggle with low overlap rates, leading to misalignments and increased computational costs as they try to find correspondences among a large number of points. This paper addresses the challenge of achieving robust and efficient point cloud registration, particularly in scenarios with limited overlap between scans.

Methods

The proposed method, LCP-4PCS (Large Common Plansets-4PCS), builds upon the 4PCS algorithm but introduces a significant modification: it prioritizes plane structures over individual points. The core steps are:

  1. Plane Segmentation: Both input point clouds are segmented into planar structures using an efficient algorithm.
  2. Iterative Plane Alignment: The algorithm iteratively applies the 4PCS method to align individual planar segments from one point cloud to segments in the other.
  3. Transformation Estimation: For each pair of aligned segments, a rigid transformation is computed.
  4. Largest Common Planeset Selection: The method identifies the transformation that results in the largest number of consistently aligned planar segments between the two point clouds. This transformation is then used to align the entire point clouds.

This approach contrasts with standard 4PCS, which aims to find the largest common pointset. By focusing on planes, LCP-4PCS is less sensitive to noise and partial overlaps, as planes provide more stable geometric features. The paper also mentions a pre-processing step to fuse small segments to prevent performance degradation.

Datasets

The authors evaluated their method on several real-world datasets:

  • Camertronix: Partial scans of building interiors captured by a Faro Focus 3D laser scanner.
  • Valentino: Data collected using stereo camera systems with integrated sensors (IMU, GNSS, laser rangefinder, infrared).
  • Charlottebürgerturm: Partial scans of a monument in Berlin, Germany, captured with stereo cameras.
  • Buro: (Likely office interior scans, based on context).
  • Flurzimmer: (Likely hallway scans, based on context).

These datasets are characterized by low overlap rates, which are specifically targeted by the LCP-4PCS method. No new benchmark datasets were introduced.

Key Insights and Ideas

  • Planes as Robust Features: The paper highlights that planar structures are more robust features for registration than individual points, especially in low-overlap scenarios. Aligning planes provides a more stable and reliable basis for transformation estimation.
  • "Divide and Rule" for Registration: The segmentation approach embodies a "divide and rule" strategy, breaking down the complex problem of aligning entire point clouds into a series of simpler alignments of planar segments.
  • Efficiency through Segmentation: By focusing on a smaller number of significant planar features rather than millions of individual points, LCP-4PCS achieves a substantial reduction in computational workload and execution time.
  • Addressing 4PCS Limitations: The research directly addresses the known limitations of 4PCS, particularly its inefficiency and potential for poor results when overlap is low or concentrated in small areas.

Key Results

  • Improved Alignment Quality: LCP-4PCS consistently produced better alignment results compared to raw 4PCS, especially on datasets with low overlap rates. This is visually demonstrated in Figure 8.
  • Significant Speedup: The method achieved approximately half the execution time of raw 4PCS across all tested datasets. For example, on the "Camertronix" dataset, execution time was reduced from 179s to 107s after segment fusion.
  • Reduced Segment Count: The segment fusion step effectively reduced the number of segments, leading to faster processing. For instance, the "Buro" dataset went from 32 segments to 9 segments after fusion.
  • Quantitative Performance: Table 2 shows that LCP-4PCS achieved a higher number of aligned segments (e.g., 8 for Charlottebürgerturm) compared to the number of aligned samples for 4PCS (e.g., 208 for Charlottebürgerturm, though this is a different metric). The execution times for LCP-4PCS were consistently lower.

Limitations and Future Work

  • Segmentation Dependency: The paper acknowledges that the performance of LCP-4PCS is highly dependent on the quality of the initial segmentation. Poor segmentation can lead to low-quality alignments.
  • High Overlap Scenarios: For datasets with near 100% overlap, raw 4PCS might converge faster because it doesn't require the initial segmentation and merging steps.
  • Future Work: The authors plan to focus on improving the segmentation step to further enhance alignment quality. They also suggest exploring better segmentation methods tailored to specific point cloud characteristics.

Summary updated

Abstract

Point cloud registration combines multiple point cloud data sets collected from different positions using the same or different devices to form a single point cloud within a single coordinate system. Point cloud registration is usually achieved through spatial transformations that align and merge multiple point clouds into a single globally consistent model. In this paper, we present a new segmentation-based approach for point cloud registration. Our method consists of extracting plane structures from point clouds and then, using the 4-Point Congruent Sets (4PCS) technique, we estimate transformations that align the plane structures. Instead of a global alignment using all the points in the dataset, our method aligns 2-point clouds using their local plane structures. This considerably reduces the data size, computational workload, and execution time. Unlike conventional methods that seek to align the largest number of common points between entities, the new method aims to align the largest number of planes. Using partial point clouds of multiple real-world scenes, we demonstrate the superiority of our method compared to raw 4PCS in terms of quality of result (QoS) and execution time. Our method requires about half the execution time of 4PCS in all the tested datasets and produces better alignment of the point clouds.
BibTeX
@article{fotsing2020large,
  author = {Cedrique Fotsing and Nafissetou Nziengam and Christophe Bobda},
  title = {{Large Common Plansets-4-Points Congruent Sets for Point Cloud Registration}},
  journal = {Isprs Int. J. geo inf},
  year = {2020},
  doi = {10.3390/ijgi9110647},
}
Figure