Abstract
Extracting planar surfaces from 3-D point clouds is essential for urban modeling and autonomous navigation. However, noise and varying point densities in airborne LiDAR and multi-view stereo (MVS) data pose significant challenges for traditional segmentation methods. This paper presents QTPS (Quasi-A-Contrario Theory-based Plane Segmentation), a novel algorithm that utilizes probabilistic "meaningfulness" to detect planes. By identifying structures statistically unlikely to occur by chance, QTPS achieves superior precision and robustness against noise compared to standard techniques like RANSAC.
Introduction
The reconstruction of 3-D city models relies heavily on the accurate extraction of flat surfaces from point clouds. While airborne LiDAR and MVS cameras provide dense spatial data, the resulting point clouds are often contaminated with noise, occlusions, and varying densities. Existing approaches—such as region growing, RANSAC-based model fitting, and feature clustering—frequently struggle with high noise levels, leading to "jittery" or inaccurately oriented planes. These methods often rely on rigid distance thresholds that fail to distinguish between true planar points and structured noise.
Methodology: The QTPS Framework
The proposed QTPS algorithm leverages Quasi-A-Contrario theory, rooted in the Helmholtz principle: a geometric structure is considered "meaningful" if its occurrence in a random noise background is statistically improbable.
Quasi-A-Contrario Theory
Instead of using a fixed distance threshold for all points, QTPS calculates the Number of False Alarms (NFA). This metric estimates the likelihood that a specific arrangement of points is a random fluke. A low NFA indicates a "meaningful" structure. The "Quasi" aspect refers to the use of a simplified white noise background model to reduce mathematical complexity.
Algorithmic Pipeline
- Multiscale Supervoxel Segmentation: The point cloud is partitioned into supervoxels to manage computational load.
- Planar Subset Extraction: The algorithm identifies the most "rigid" subset of points within each supervoxel. Subsets with a low NFA are designated as Fine Planar Supervoxels (FPS).
- FPS Optimization: These high-precision seeds are refined by incorporating nearby unclassified points that statistically belong to the plane.
- Region Growing: Finally, adjacent FPS units with similar orientations are merged to form continuous large-scale planes, such as building roofs or facades.
Experimental Evaluation
The performance of QTPS was validated using simulated data and three real-world datasets: Dublin City Center (high noise ALS), Ningbo (sparse ALS), and Dortmund (MVS).
Key Results
- Precision: In the Dublin dataset, QTPS achieved a Root Mean Square Error (RMSE) of 0.0183m, significantly outperforming Efficient RANSAC (0.0286m) and Octree-based Region Growing (0.0446m).
- Robustness: Simulated tests demonstrated that QTPS maintains high accuracy even with a noise ratio as high as 75%.
- Detail Preservation: The method successfully preserved small architectural features like eaves and windowsills, which are often lost in traditional smoothing processes.
Conclusion
QTPS provides a robust and precise solution for 3-D plane segmentation by shifting the focus from hard distance thresholds to statistical meaningfulness. While the algorithm can occasionally detect "pseudo-planes" in extremely structured noise and may be slower than simpler methods, its ability to handle high-noise environments and preserve fine details makes it a powerful tool for building reconstruction. Future work will focus on integrating topological analysis to further refine the logical structure of the generated 3-D models.
Summary updated