Classification of sensor independent point cloud data of building objects using random forests

TL;DR: A generic approach to automatically identify structural elements for the purposes of Scan-to-BIM by taking a set of planar primitives that are pre-segmented from the point cloud.

Abstract: The Architectural, Engineering and Construction (AEC) industry is looking to integrate Building Information Modeling (BIM) for existing buildings. Currently these as-built models are created manually, which is time-consuming. An important step in the automated Scan-to-BIM procedure is the interpretation and classification of point cloud data. This is computationally challenging due to the sheer size of point cloud data for an entire building. Additionally, the variety of objects makes classification problematic. Existing methods integrate prior knowledge from the sensors or environment to improve the results. However, these approaches are therefore often case specific and thus have limited applicability. The goal of this research is to provide a method that is independent of any sensor or scene within a building environment. Furthermore, our method processes the entire building simultaneously, resulting in more distinct local and contextual features. This paper presents a generic approach to automatically identify structural elements for the purposes of Scan-to-BIM. More specifically, a Random Forests classifier is employed for the classification of the floors, ceilings, roofs, walls and beams. As input, our algorithm takes a set of planar primitives that are pre-segmented from the point cloud. This significantly reduces the data while maintaining accuracy. Both contextual and geometric features are used to describe the observed patches. The algorithm is evaluated using realistic data for a wide variety of existing buildings including houses, school facilities, a factory, a castle and a church. The experiments prove that the proposed algorithm is capable of properly labeling 87% of the structural elements with an average precision of 85% in highly cluttered environments without the support of the sensors position. In future work, the classified patches will be processed by class-specific reconstruction algorithms to create BIM geometry.

Summary

1. Introduction

• The implementation of Building Information Modelling (BIM) for existing buildings is gaining popularity.
• Experiencing the advantage of BIM for new constructions, the industry now looks to implement as-built BIM.
• These as-built models store an immense amount of information about a building at the varying stages of the construction’s life cycle [1].
• More specifically, structural elements such as floors, ceilings, roofs, walls and beams are automatically identified in existing structures.
• 35 In Section 4 the methodology is presented.

2. Background

• The procedure of converting point cloud data to BIM geometry is referred40 to as Scan-to-BIM.
• Second, each cluster is provided with a class label.
• Examples of local geometric features are the area, surface dimensions and orientation.
• Heuristic models are based on user defined rules in a certain structure.
• Alternatively, machine learning algorithms are60 employed such as Discriminant Analysis (DA), Decision Trees, Support Vector Machines (SVM), Neural Networks (NN), Probabilistic Graphical Models (PGM), etc. [8, 12, 13, 16, 17, 18, 19].

3. Related Work

• Semantic labelling of building geometry has been a major research topic over the last decade.
• Classification algorithms are often linked to a certain sensor set up or point cloud format [8, 15, 34, 35].
• A classification procedure is presented for the labelling of structural objects in built environments.
• The vertical topology feature Dz(u ∈ si ,v ∈ sj ) evaluates the vertical distance between the observed surface si and the large horizontal surfaces sj within its vicinity (Fig. 3 b).
• The proximity feature captures the repetitivity of certain object configurations.

4.2. Model formulation

• Each tree consists of a series of binary splits that separate the input variables.
• The Random Forests model is trained using leave-p-out cross validation.
• This intuitive procedural programming platform allows for flexible data processing and evaluation.
• The classified patches are exported to the Rhinoceros model space for185 validation and further processing.

5. Experiments

• 10 structures including houses, offices, industrial buildings and churches were used for training and testing (Fig 5).
• The test sites were acquired under realistic conditions including clutter, occlusions, traffic, etc.
• Ghz with 4 cores and 4 hyperthreads and 32GB RAM.
• Over 90,000 surfaces were computed for the projects.
• All 17 predictors from table 1 were considered for the classification of the observations.

5.1. Performance

• The classification results are depicted in the confusion matrices in Fig.7.
• This is very accurate given the large variety of buildings and objects that were evaluated.
• Increased confusion rates are observed between the walls and clutter classes as well as the ceiling and roof classes.
• This is due to the fact that several data sets do not have roofs making the top ceilings harder to230 interpret (Fig.8b).
• Several misclassifications are due to their sensor independent approach.

5.2. Comparison

• The authors compared the results of the Random Forests classifier with other common machine learning methods.
• Table 2 depicts the results of the model performance for K-Nearest Neighbours (KNN), a multiceptron Neural Network (NN), Support Vector Machines (SVM) and boosted decision trees.
• All models240 were tested with the same predictors and data as the proposed model.
• This proves that the used predictors are both distinct and robust for the detection of structural elements in cluttered and noisy environments.
• Since their approach focusses on post-processing applications, the training time is of lesser concern.

6. Discussion and Conclusion

• More specifically, the data is pre-segmented and processed by machine learning algorithms to label the floors, ceilings, roofs, beams, walls and clutter in noisy and occluded environ-255 ments.
• This allows for the processing of larger data sets and provides additional features.
• Some classes underperform due to the large variance in feature values within the265 class.
• This will enhance the current classification and allows for the processing of non-planar classes such as cylindrical beams280 and pipes as well as furniture.

Figures

Figure 4: Representation topology predictors: Topology small surfaces (Left), evaluation topology nearby significantly large horizontal surfaces (Right).

Figure 7: Classification results Random Forests classifier: Recall (Left), Precision (Right). The percentage defined under the ’True Class’ is the amount of available test data.

Figure 3: Visualisation of class specific neighbours: Conventional nearest neighbours (a), significantly large nearby surfaces (b), significantly large horizontal surfaces above/below (c) and nearby small surfaces (d).

Figure 6: Example test data with ground truth: Laboratory (a), classroom (b), chemical plant (c), flat roof structure (d), heritage roof beam structure (e) and heritage room (f). Red=floors, purple=ceilings, blue=roofs, green=walls, yellow=beams and grey=clutter.

Figure 2: Pre-segmentation: Registered point cloud (Left) and the resulting planar patches (Right).

Figure 1: Example point clouds of structures used during testing: Chemical facility (a), house (b), multi-storey school building (c) and a church (d).

Table 2: Averaged classification results of Random Forests compared to K-Nearest Neighbours, multiceptron Neural Network, Linear Support Vectors Machines and Boosted Decision Trees.

Figure 5: Example test cases with reconstructed surfaces: House 4000 surfaces (a), Campus 30000 surfaces (b), Office 26000 surfaces (c), Row house 1800 surfaces (d), Multi-storey office 3000 surfaces (e), heritage site 10000 (f), church 6500 surfaces (g) and chemical plant 15000 surfaces(h).

Table 1: List of all features calculated for each observed patch

Figure 8: Several misclassification examples. Large closets, doors and machinery labelled as walls (a,c and d), unobstructed ceilings labelled as roofs (b), near vertical ceilings labelled as walls (e) and unobstructed ceilings labelled as floors (f).

Frequently asked questions

Q1. What contributions have the authors mentioned in the paper "Classification of sensor independent point cloud data of building objects using random forests" ?

The goal of this research is to provide a method that is sensor independent and labels entire buildings at once. This paper presents a method to automatically identify structural elements for the purposes of Scan-to-BIM. The experiments prove that the proposed algorithm is capable of labelling structural elements with reported precisions of 85 % and 87 % recall in highly cluttered environments. 

Q2. What are the future works in "Classification of sensor independent point cloud data of building objects using random forests" ?

In future work, the method will be investigated further to improve the labelling performance. Also, research will be performed towards the integration of probabilistic graphical models to increase the methods perfor- mance.