Complete residential urban area reconstruction from dense aerial LiDAR point clouds

Qian-Yi Zhou, Ulrich NeumannGraph. Model 2013 · DOI

TL;DR

This paper presents a novel system for automatically reconstructing 3D urban models of residential areas from aerial LiDAR scans. It tackles the challenge of rich vegetation by introducing a classification method that leverages the "2.5D characteristic" of buildings (opaque roofs and vertical walls) to distinguish them from trees, enabling the creation of hybrid models that accurately represent both structures.

What Problem Does This Paper Address?

Traditional urban modeling research has largely focused on downtown areas, where buildings are the primary focus and vegetation is often treated as an obstruction to be removed. However, residential areas present a different challenge: vegetation (trees) is a significant component of the urban landscape and needs to be modeled alongside buildings. Furthermore, dense LiDAR scans can capture detailed geometry of tree crowns, which can exhibit similar local geometric features (like planarity) to building rooftops, making traditional classification methods unreliable. This paper aims to bridge this gap by developing a robust system that can accurately classify and reconstruct both buildings and trees in vegetated residential areas.

Methods

The core of the proposed system is a novel classification algorithm that leverages the "2.5D characteristic" of buildings. This characteristic, defined as buildings being composed of opaque skyward roofs and vertical walls, means that laser beams cannot penetrate the structure, and the rooftop or ground is the lowest visible surface at a given (x,y) location. Trees, on the other hand, have a more complex structure with surfaces both within and under the crown, and do not adhere to this strict 2.5D property.

The classification process involves:

  1. Discretization and Layer Fragment Segmentation: The point cloud is discretized into a 2D grid, and within each grid cell, points are segmented into multiple "layer fragments" using local distance-based region growing.
  2. 2.5D Characteristic Application: The algorithm assigns the lowest layer fragment in a cell as non-tree (building or ground) because it represents the lowest visible surface. Other layer fragments are classified as trees. This is formulated as a data energy term in an energy minimization framework.
  3. Energy Minimization: A Markov Random Field (MRF) is constructed to further discriminate between buildings and ground, incorporating elevation differences and smooth energy terms. This MRF is solved using a graph-cut method to label each layer fragment.
  4. Modeling: Once classified, building models are created using a 2.5D dual contouring method, with a novel hole-filling step to address occlusions caused by trees. Tree models are generated by fitting inscribed octahedrons to local leaf shapes around tree points, derived from Principal Component Analysis of neighboring points. Ground models are created by rasterizing ground points into a Digital Surface Model (DSM).

Datasets

The paper evaluates the system on three distinct datasets:

  • Atlanta area #1
  • Atlanta area #2
  • Denver area

These datasets vary in point density and resolution, demonstrating the system's robustness. The authors do not introduce new benchmark datasets but rather use existing aerial LiDAR scans.

Key Insights and Ideas

  • The "2.5D Characteristic" as a Discriminator: The key insight is that the inherent geometric property of buildings (opaque, skyward-facing surfaces) provides a robust way to distinguish them from trees, even when local geometric features are similar. This is a significant departure from methods that solely rely on local geometry.
  • Trees as a Feature, Not an Obstacle: Instead of removing trees, the system actively models them, creating a more complete and realistic representation of residential areas.
  • Hybrid Modeling: The combination of 2.5D building models and leaf-based tree models results in a hybrid representation that captures the complexity of residential urban environments.
  • Robustness to Occlusion: The proposed hole-filling step for building reconstruction effectively handles occlusions caused by trees, which is a common problem in aerial LiDAR data of vegetated areas.

Key Results

  • Classification Accuracy: The classification approach, leveraging the 2.5D characteristic, successfully classifies points into trees, buildings, and ground, as demonstrated in Figure 4.
  • Reconstruction Quality: The system generates visually appealing and accurate hybrid urban models.
  • Quantitative Evaluation: In experiments, no false negatives were found for building detection, meaning all building structures were successfully detected and reconstructed. False positives (e.g., small building-like trees, misclassified ground) were present but generally low, with error rates calculated based on building models.
  • Efficiency: The system is efficient, with total reconstruction times for the tested datasets ranging from 23 to 109 seconds on a consumer-level laptop. For a 5.5M point dataset, it reconstructs 56K triangles for buildings and 53K octahedrons for trees in under two minutes.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that small building-like trees and incorrectly classified ground components can lead to false positives.
  • Buildings under construction, lacking roofs, are currently misclassified as trees.
  • The quality measurement for ground truth classification is challenging due to similar local geometry properties and spatial proximity of buildings and trees, leading to the adoption of a building error rate.
  • Future Work:
    • Expanding Urban Element Categories: The authors suggest exploring other urban element categories, such as walls of buildings under construction, which are currently misclassified.
  • Data Fusion: Integrating aerial LiDAR with other data sources like aerial imagery, ground LiDAR, and street view images is proposed to create more realistic models by combining geometry and visual information.

Summary updated

BibTeX
@article{zhou2013complete,
  author = {Qian-Yi Zhou and Ulrich Neumann},
  title = {{Complete residential urban area reconstruction from dense aerial LiDAR point clouds}},
  journal = {Graph. Model.},
  year = {2013},
  doi = {10.1016/j.gmod.2012.09.001},
}
Figure