Automatic Creation of Semantically Rich 3D Building Models from Laser Scanner Data

TL;DR: A method to automatically convert the raw 3D point data from a laser scanner positioned at multiple locations throughout a facility into a compact, semantically rich information model that is capable of identifying and modeling the main visible structural components of an indoor environment despite the presence of significant clutter and occlusion.

About: This article is published in Automation in Construction.The article was published on 2013-05-01 and is currently open access. It has received 576 citations till now. The article focuses on the topics: Laser scanning & Point cloud.

Summary

1. INTRODUCTION

• In the Architecture, Engineering, and Construction (AEC) industry, semantically rich 3D models are increasingly used throughout a building’s lifecycle, from the design phase, through construction, and into the facility management phase.
• Currently, as-is BIMs are created through a manual process, typically using data from laser scanners as input [1].
• Applied to all the rooms in a building, their method will automatically produce a compact, semantically rich, 3D model which, while not strictly a BIM in the traditional sense, contains the geometric and identity information that substantially makes up the BIM.
• The authors algorithm addresses the challenges of clutter and occlusion by explicitly reasoning about them throughout the process.
• These two phases are described in more detail in the next two sections.

2. CONTEXT-BASED MODELING

• Distinguishing between different types of objects and between clutter and non-clutter can be difficult or impossible if those objects are seen only in isolation.
• The more easily recognized structures can provide the scaffolding that enables the recognition of other, more challenging instances.
• This phase of the algorithm consists of four steps .
• Patches are found using a region growing algorithm to connect nearby points that have similar surface normals and that are welldescribed by a planar model.
• The result of this process is a compact model of the walls, floor, and ceiling of a room, with each patch labeled according to its type.

3. DETAILED SURFACE MODELING

• The surfaces produced by the first phase represent idealized surfaces that are perfectly planar and unoccluded and with no openings.
• By tracing a line from the origin of the laser scanner to a detected measurement, a pixel is classified as empty if the ray passes through the plane of the patch, as occupied if the ray stops approximately at the plane, and occluded if the ray stops before reaching the plane.
• Detecting openings in unoccluded surfaces can be achieved by analyzing the data density and classifying low density areas as openings.
• The authors use a number of features as input to the SVM, including opening width, height, and distance from the boundaries of the patch (sides, and top, and bottom).
• This step is not strictly necessary, but it improves the visualization of the results .

4. EXPERIMENTAL RESULTS

• The authors conducted experiments using data from a building that was manually modeled by a professional laser scanning service provider.
• The patch detection resulted in 389 patches in the interior of the building.
• The highest confusion was between walls and clutter.
• The authors experiments suggest that the context aspect of their algorithm improves recognition performance by about 6% and that the most useful contextual features are coplanarity and orthogonality.
• Figure 5 shows some example results, and Figure 6 shows one floor of the entire reconstructed model.

Figures

Figure 5. Example results of the detailed surface modeling algorithm for entire rooms. a) The room shown in Figure 2. Walls, ceilings, and floors are outlined in blue, and windows and doors are outlined in red. b) Another room. c) An example where the opening detection failed due to an arched window, which violates the assumptions of the algorithm.

Figure 6. Results for all rooms on the second floor. The color scheme is the same as in Figure 5a.

Figure 1. As-is BIM creation. a) The input data from a single scanner location – reflectance image (top), range image (bottom), and close up of highlighted region (right). b) The point data is aligned and combined into a cloud of points. Data from four scans was combined to form this point cloud. c) Geometric primitives are modeled to form the BIM. This BIM was created manually by a professional service provider.

Figure 2. The context-based modeling algorithm consists of four steps, shown for the point cloud from Figure 1: a) voxelization; b) patch detection; c) patch classification; d) patch intersection and clutter removal.

Figure 3. The detailed surface modeling algorithm, shown for one wall using the results of Figure 2: a) occlusion labeling; b) iterative region growing; c) opening detection; d) occlusion reconstruction.

Figure 4. Examples of the context-based modeling algorithm for some other rooms. The right example is a top view of a bathroom, in which the stalls and some large cabinets were all successfully distinguished from the walls.

Frequently asked questions

Q1. What are the contributions in "Automatic creation of semantically rich 3d building models from laser scanner data" ?

This paper presents a method to automatically convert the raw 3D point data from a laser scanner positioned at multiple locations throughout a building into a compact, semantically rich model. Then, the authors perform a detailed analysis of the recognized surfaces to locate windows and doorways. The authors evaluated the method on a large, highly cluttered data set of a building with forty separate rooms yielding promising results. 

Q2. What are the useful features of the context-based modeling algorithm?

Their experiments suggest that the context aspect of their algorithm improves recognition performance by about 6% and that the most useful contextual features are coplanarity and orthogonality. 

Q3. What is the first phase of the BIM process?

In the first phase, planar patches are extracted from the point cloud and a context-based machine learning algorithm is used to label the patches as wall, ceiling, floor, or clutter. 

Q4. What is the purpose of the detailed surface modeling phase of the algorithm?

The detailed surface modeling phase of the algorithm operates on each planar patch produced by the contextbased modeling process, detecting the occluded regions and regions within openings in the surface. 

Q5. What is the purpose of the algorithm?

A learning algorithm is used to encode the characteristics of opening shape and location, which allows the algorithm to infer the shape of an opening even when it is partially occluded. 

Q6. What is the common problem in the BIM pipeline?

The authors are currently working on completing the points-to-BIM pipeline by implementing an automated method to convert the surface-based representation produced by their algorithm into a volumetric representation that is commonly used for BIMs. 

Q7. What is the way to detect openings in unoccluded surfaces?

Detecting openings in unoccluded surfaces can be achieved by analyzing the data density and classifying low density areas as openings. 

Q8. What is the purpose of the classifier?

The classifier uses local features computed on each patch in isolation as well as features describing the relationship between each patch and its nearest neighbors. 

Q9. What are the main purposes of building information models?

These models, which are generally known as building information models (BIMs), are used for many purposes, including planning and visualization during the design phase, detection of mistakes made during construction, and simulation and space planning during the management phase. 

Q10. What is the result of this process?

The result of this process is a compact model of the walls, floor, and ceiling of a room, with each patch labeled according to its type. 

Q11. What are the challenges of building modeling algorithms?

Building modeling algorithms are frequently demonstrated on simple examples like hallways that are devoid of furniture or other objects that would obscure the surfaces to be modeled.