AD HOC MODEL GENERATION USING MULTISCALE LIDAR DATA FROM A GEOSPATIAL DATABASE

M. Gordon, Björn Borgmann, J. Gehrung, Marcus Hebel, Michael Arens2015 · DOI

TL;DR

This paper presents a novel approach for generating 3D models from multi-scale LiDAR data by intelligently storing and querying point clouds in a geospatial database. By fusing data from different sources (airborne, mobile, terrestrial) with varying accuracies, their octree-based method creates comprehensive and precise models suitable for real-time visualization, achieving significant data compression.

What Problem Does This Paper Address?

The proliferation of affordable laser scanning technology has led to an explosion of 3D point cloud data. Visualizing and processing this massive amount of data in real-time for applications like topographic surveying and mapping is a significant challenge. Traditional methods struggle with the sheer volume, and efficiently storing and retrieving specific data subsets based on spatial criteria is also problematic. This paper tackles the problem of creating detailed 3D models from diverse LiDAR sources by addressing both efficient data management and intelligent model generation.

Methods

The core of the proposed solution lies in two main components:

  • Geospatial Database for Data Management: They developed a custom geospatial database to store 3D point clouds along with crucial metadata. This metadata includes geolocation (latitude/longitude), coverage (bounding box), and importantly, the standard deviation of the sensor's positional accuracy. This allows for efficient querying of data based on regions of interest (ROI) and other spatial parameters, as well as filtering by data quality.
  • Octree-Based Model Generation with Data Fusion: The model generation process uses an octree data structure to handle the hierarchical organization of space. For each octree cell, the goal is to fit a single plane. This is achieved using a RANSAC (Random Sample Consensus) algorithm. A key innovation is the data fusion aspect:
  • Multi-scale Integration: Lower-density airborne LiDAR data is used to supplement higher-density mobile and terrestrial data, improving overall coverage.
  • Accuracy-Aware Plane Fitting: The standard deviation of sensor accuracy is used to define a threshold for point-to-plane distance, allowing points from different sources with varying precisions to be integrated.
  • Weighted Principal Component Analysis: The standard deviation also serves as a weighting factor in principal component analysis, further enhancing the fusion of data with different accuracies.
  • Point Handling: Points are processed sequentially. If a cell already has a plane, new points are classified as supporting or contradicting. If too many contradicting points exist, a new plane fit is attempted or the cell is subdivided. For empty cells, points are collected until a plane fit is feasible, otherwise, the cell is subdivided.
  • Uniformity Test: A test based on principal component analysis ensures that the fitted plane is a good representative of the points within the cell, preventing the loss of detail. If the distribution is not uniform, the cell is further subdivided.

Datasets

The paper utilizes a combination of three types of LiDAR data for their experiments:

  • Airborne Laser Scanning (ALS): Data collected using a RIEGL LMS-Q560 scanner, providing broad coverage but with lower precision (standard deviation of 0.3 m).
  • Mobile Laser Scanning (MLS): Data acquired using Velodyne HDL-64E scanners mounted on a vehicle, offering high point density and accuracy (standard deviation of ~1.7 cm).
  • Terrestrial Laser Scanning (TLS): Data from a Zoller+Fröhlich IMAGER 5003 scanner, providing the highest accuracy (standard deviation of <0.5 cm) but with slower acquisition.

The data was collected from an urban area around the Fraunhofer IOSB in Ettlingen, Germany.

Key Insights and Ideas

  • Database as a Foundation for 3D Data Processing: Moving beyond simple storage, the geospatial database acts as an intelligent intermediary, enabling efficient retrieval and filtering of point cloud data based on spatial criteria and accuracy, which is crucial for complex modeling tasks.
  • Synergistic Fusion of Multi-Scale LiDAR: The paper demonstrates that combining data from different LiDAR sources, each with its own strengths and weaknesses (e.g., ALS for coverage, TLS for detail), can lead to a model that surpasses the quality of any single source.
  • Octree for Adaptive Detail and Compression: The octree structure naturally handles varying levels of detail in the data, allowing for fine-grained representation in dense areas and broader approximations where data is sparse. This adaptive nature is key to achieving high data compression ratios.
  • Accuracy-Aware Integration is Paramount: Simply merging point clouds is insufficient. The method's success hinges on explicitly considering the accuracy of each data source during the model generation process, ensuring that less precise data doesn't degrade the overall model quality.

Key Results

  • Comprehensive Model Generation: The approach successfully generates polygon models that represent planar surfaces in the urban environment, while also preserving points that cannot be represented by planes (e.g., vegetation).
  • Data Compression: The method achieves significant data compression ratios, with figures around 90% reported, making real-time visualization feasible.
  • Improved Completeness and Precision: The fused models exceed the completeness and precision of individual data sources. For instance, incorporating ALS data significantly improves the coverage of building roofs in areas where only sparse airborne data was previously available.
  • Parameter Sensitivity: The paper explores the impact of various parameters on model generation and data compression, highlighting trade-offs between detail preservation and compression efficiency. For example, a higher sample coverage factor leads to better compression but can result in the loss of small details.

Limitations and Future Work

  • Vegetation Representation: The current approach struggles with representing vegetation, particularly trees, as they do not typically consist of planar surfaces. This leads to chaotic octree cells and suboptimal compression in these areas.
  • TLS Georeferencing: The paper assumes data is georeferenced, but notes that TLS data often requires automatic registration with other sensor data (MLS/ALS), which needs to account for differing precisions.
  • Low ALS Density: The low density of ALS data can pose challenges for model creation, leading to gaps in the generated model.
  • Future Work: The authors suggest exploring more complex database queries beyond simple ROIs, such as retrieving data along a specific path for navigation or visualization. They also highlight the need for robust methods to register TLS data with other sensor types, considering their varying accuracies. Further research into handling non-planar objects like vegetation is also implied.

Summary updated

Abstract

Due to the spread of economically priced laser scanning technology nowadays, especially in the field of topographic surveying and mapping, ever-growing amounts of data need to be handled. Depending on the requirements of the specific application, airborne, mobile or terrestrial laser scanners are commonly used. Since visualizing this flood of data is not feasible with classical approaches like raw point cloud rendering, real time decision making requires sophisticated solutions. In addition, the efficient storage and recovery of 3D measurements is a challenging task. Therefore we propose an approach for the intelligent storage of 3D point clouds using a spatial database. For a given region of interest, the database is queried for the data available. All resulting point clouds are fused in a model generation process, utilizing the fact that low density airborne measurements could be used to supplement higher density mobile or terrestrial laser scans. The octree based modeling approach divides and subdivides the world into cells of varying size and fits one plane per cell, once a specified amount of points is present. The resulting model exceeds the completeness and precision of every single data source and enables for real time visualization. This is especially supported by data compression ratios of about 90%.
BibTeX
@inproceedings{gordon2015ad,
  author = {M. Gordon and Björn Borgmann and J. Gehrung and Marcus Hebel and Michael Arens},
  title = {{AD HOC MODEL GENERATION USING MULTISCALE LIDAR DATA FROM A GEOSPATIAL DATABASE}},
  booktitle = {˜The},
  year = {2015},
  doi = {10.5194/isprsarchives-xl-3-w3-535-2015},
}
Figure