Textured As-Is BIM via GIS-informed Point Cloud Segmentation

Due to the ailing state of the German railways, an increased focus on digitisation is directed recently to maintain, upgrade and expand the existing railway network and prepare it for the future [1, 2]. Digitisation of the standing network is a very costly and time-consuming endeavour, especially in Germany, where over 33,000 km of railway exist, with almost 2,000 km of railway tracks, 200 bridges and 18,000 points and crossings renovated and upgraded in 2023 alone [3]. This is even more pronounced in the early stages of the larger construction projects, upon which status-quo data (e.g., plans, drawings, or surveys) depends, that is often incomplete, exists in non-digital formats or could be missing altogether [4]. However, this data still needs to be processed for integration into inter-compatible working methods based on Building Information Modelling (BIM) [5, 6, 7].

The time-frame of such processes from a project’s initiation to the finishing handover often spans over decades as evident from some of the German Reunification Railway Transport Projects number 8 and 9 for the expansion and new construction of the routes Nuremberg–Erfurt–Halle/Leipzig–Berlin, and along the route between Leipzig –Dresden respectively [8], where the acquisition of up-to-date surveying data of the location is crucial to ensure correct planning and informed decision-making. Especially through the very inceptive phase of the project, when there is often little to no allocated funds to support the decisions needed to be made that are in turn time-varying on shifting requirements, demand, possible project variants, extents, costs, regulations and permits, etc [9].

A patchy data basis, however incomplete it might be, could be improved upon with registered surveying patches for regions of interest using commercial aerial Light Detection and Ranging (LiDAR) scanning services, through photorealistic 3D tiles from the latest Google Maps 3D immersive view [10] or other similar alternatives. Yet, surveying services are often very expensive, rendering their costs a hindrance to utilise, as they are neither always available, easy to acquire at remote locations nor guaranteed to be up to date over time with changes of the status-quo on ground. Thus, a more practical and less costly approach, that includes openly accessible geographic and location information free of charge provided by online web mapping platforms and Geographic Information Systems (GIS), could be of benefit. The open geodata published by official surveying authorities and available for free could be an economical alternative source to acquire a relatively up-to-date data basis that can be processed and utilised to generate as-is BIM models.

The latest strides in the research field of intelligent recognition of Point Cloud Data (PCD) have been established in state of the art systems for autonomous driving, navigation, flight planning for drones, and aerial scans’ segmentation, that could be utilised for the semantic segmentation of point clouds generated from available geodata. This would help classify further objects of interest that are unlabelled in the raw data or difficult to attain otherwise to derive further insightful knowledge and semantic information that assist in the identification of interdependencies and relations between labelled classes and project requirements defined by official regulations, public authorities, planners and other concerned stakeholders.

Furthermore, the various types of data and information formats, compounded by limitations to incompatible software and hardware used by each actor, emphasises the benefits to common adoption of freely and widely available open data formats, which are information-lossless, convertible and inter-compatible with technical hardware and software resources of all involved stakeholders. This early phase always requires extensive data exchange between various stakeholders, that highlights the importance of adopting big open BIM to facilitate interoperability and reduce the incompatibility of different data formats separately used by each party involved.

Such reconstruction of the status-quo, as relatively accurate BIM-ready as-is railway models they are, using freely available PCD (i.e., Scan2BIM) suffices to propel information based decision-making forward in early planning phases immensely with minimum cost. A hybrid approach that incorporates not only PCD, but also geospatial and a mix of other free publicly accessible data may help overcome common obstacles arising from insufficient data basis.

Hence, this study considers the applicability of state of the art methods available today to find practical solutions to the main question of how to optimally combine the wealth of information available in open geodata, including the German official topographic and property cadastre information systems (i.e., photos, point clouds, property and land use shapes and attributes) through semantic segmentation of point clouds. The focus lies on the railway infrastructure domain, with BIM models in the open IFC format, which paves the way for a very economical workflow to reconstruct an as-is information model well suited for BIM-based planning in early project phases. It focuses on the transfer of knowledge attained from semantic PCD segmentation through surface reconstructed geometry and derived semantics and properties into big open BIM. We hypothesise that the inclusion of GIS data may help overcome technical and financial limitations to insufficient data basis. This research further expands on previous work by the authors [11]. For this purpose, we rely solely on freely available data, using only open-source tools and convert our results into the open BIM IFC data format.

A modified version of the deep neural network (DNN) architecture 2DPASS has been used to train a model capable of semantically segmenting buildings, vegetation, water bodies, roads, and railway trackbeds. The used architecture relies in the training phase both on 3D features from a dataset of processed and semi-automatically annotated LiDAR scans as well as 2D features from coloured orthophotos. The original dataset is acquired from freely available LiDAR scans, digital orthophotos, ATKIS shape files from the geoportal websites of the Free States of Thuringia and Saxony. The authors of this study attempt to make the most use of free and open data sources, software, and information formats for pre- and post-processing of point cloud data, semantic segmentation, post processing of the inferred segmentation results with 3D reconstruction and modelling of railway related contexts.

The contribution of this work links to formerly published literature about Scan2BIM approaches in railway [12, 13, 14, 15, 16]. The proof of concept to our proposals is demonstrated on two case studies. The first case study demonstrates the potentials for cost savings in early planning phases of railway projects by estimating horizontal and vertical alignments from the inferred semantic segmentation of the predicted railway trackbeds in GIS-compatible data format (i.e., GeoJSON) that could be later used as input for automatically generating and IfcAlignment or querying multi- or bi-temporal differences along a buffer zone of the alignment.

The second case study demonstrates the texturing potentials of the resulting IFC files and builds on former studies by [17, 18, 19], and expands it further as no attempt has been found in the body of literature so far to generate UV-mapped textured meshes of irregular 3D shapes in IFC automatically, whereas the aforementioned studies successfully embedded a texture image onto a 2D flat surface in 3D space with manual UV-mapping.

Point cloud acquisition has become more ubiquitous due to advancements in latest depth and light capturing sensors, that could be mass-produced within a wide range of price, quality and application spanning from LiDAR scans in engineering surveying to those fitted on drones or the miniaturised variants in smartphones and extended reality (XR) gadgets. This wide adoption has driven progress on semantic segmentation of point clouds to develop highly sophisticated Deep Learning (DL) models able to classify large amounts of points automatically on graphical and computer processing units for a wide range of applications ranging from early phase planning of projects, scene understanding for autonomous driving to construction monitoring and inspection, XR applications and beyond.

Texture utilising methods for point cloud segmentation involve a collection of varying approaches. One focuses on fusing representations from points, voxels, and/or projection images within different branches of the network’s architecture. Tang et al. combined point-wise MLPs in each sparse convolution block to learn a point-voxel representation [20] and relied on Neural Architecture Search to reach an optimally efficient design . Xu et al. proposed a range-point-voxel fusion network (i.e., RPVNet) to utilise information from the three representations [21]. However, texture features and semantics from photos were underutilised or discarded altogether by focusing only on colourless LiDAR point clouds.

Another approach fuses inputs from multiple sensors leveraging benefits of both camera and LiDAR [22, 23, 24]. El Madawy et al. converted RGB images to a polar-grid mapping representation and designed early and mid-level fusion architectures [22]. Point Painting exploited the segmentation logits of images [25] and projected them into LiDAR space by spherical [26] or bird’s-eye projection [27] for LiDAR network performance improvement, whereas Zhuang et al. exploited a collaborative fusion of two modalities in camera coordinates [28]. Yet, the paired multi-modality data requires multi-sensor inputs in both training and inference and is more computationally intensive and unavailable in practical applications.

A third approach involves knowledge distillation, which is a technique used to transfer knowledge learnt by compressing a large teacher network to a smaller student [29]. Further improvements to knowledge transferring were achieved using different methods of matching feature representations [30, 31]. For instance, aligning attention maps [32] and Jacobean matrices [33] were independently applied. This technique has been applied successfully to transfer priors across different modalities by using additional images in the training phase and to improve performance at inference [34, 35, 36, 37, 38]. For instance, by either inflating kernels of 2D convolution into 3D [21], introducing a 2D pixel-to-point assisted pre-training [39] or by utilising a teacher-student framework [40]. Furthermore, 2DPASS leveraged auxiliary modal fusion to transfer 2D knowledge through multi-scale fusion-to-single knowledge distillation, which took care of the modal-specific knowledge and demonstrated generality, flexibility and effectiveness when compared to other fusion based methods [41].

The so-called Scan2BIM approach means the (automated) acquisition of PCD and the subsequent creation of BIM models from it  [42]. Many works deal with the topic of applying Scan2BIM to railway infrastructure. This subsection presents an overview of the related research.

Ariyachandra et al. [13] described a method for creating large-scale geometric information models (GIM) of rail infrastructure using automated segmentation of point cloud data (PCD). The geometry of the segmented rails and track bed was then reconstructed resulting in IFC files of the respective objects. The authors distinguished between a distinct segmentation of rails and track beds. In their earlier publications [43], they focused on the identification of railway masts of double-track railways from LiDAR PCD by taking railway design rules strongly into account and therefore renouncing neighbourhood structures, scanning geometry and the intensity of input data. After removing the adjacent vegetation, a track corridor remains from which the masts were then extracted, which are parallel to the global Z-axis. The masts were then differentiated from other pole-like objects defining inner and outer boxes around the masts to identify outliers. The proposed method reached high detection and precision rates, and included eventually the 3D mast models in IFC.

Yang [44] extracted rail beds and rails from mobile laser scanning (MLS) point cloud. They were able to detect different rail segments (where one alignment ends and the other begins, e.g., points) and reached an overall detection accuracy of more than 95% in completeness and quality and 99.78% correctness. The applied method used cross-sections and so called “scanning lines” to identify the ballast bed of the rail and with the help of the specific geometric angle of the ballast bed, and was able to identify the course of the railway and to identify objects that did not belong to the railway itself. To avoid scanning the whole point cloud, the areas of interest have been segmented in the point cloud. In order to quantify the accuracy of the result, rail lines were manually digitised as lines and with a buffer zone of 15 cm compared to the extracted railway lines from the point cloud.

Cheng et al. [45] used coloured point clouds (range accuracy $\pm 2~{}mm$) in order to first segment them into (single-track) railway bed and remaining points. The tunnel cross-sections were extracted and the parameters derived. Ten different classes were identified, including rail head and railway bed. To estimate the horizontal and vertical alignments of the railway, a method of 3D local cylindrical neighbourhood was used. The presented method for classifying point cloud reached close to 100 % accuracy. The authors achieved to create a parametric BIM model from the derived objects, yet it was not converted into an open product data model.

Eickeler and Borrmann [14] presented a three-stage concept for the creation of a railway specific dataset that allows a higher grade of automation in railway asset detection. The paper focused on videos and images, not on PCD derived from LiDAR scans. The three stages consisted of 1. Simple Class Annotation (SCA), 2. classification by domain knowledge and 3. creation a full asset model (optimisation).

Cserép et al. [15] used coloured LiDAR PCD from a vehicle-mounted scanner (mobile mapping system; 60 km/h). The PCD was fragmented and then used for cable and rail recognition. To remove vegetation within the PCD along the railway tracks, contour detection was used. The railway was identified using contour finding and Hough transformation. The paper compared three different algorithms for detection of the railways. First, the trackbed was detected taking into account the point density and the most common height in a defined area. Second, the PCD was reduced to a 2D digital elevation model (image) and the 2D Hough transformation used in order to identify rail pair from the resulting 2D projected PCD. With this methodology of creating as-is models, it was hard to detect cables being placed underneath each other, but the standardised track gauge could be taken into account. The 3D line of the railway was detected using Hough transformation. Third, the region growing algorithm with the trajectory of the identified cables was used. The trajectory of the power cables was calculated with the Random Sample Consensus (RANSAC) algorithm. All three approaches have been compared. For the cable objects, the Region growing algorithm showed the best accuracy and operated the fastest.

The work of Soilán et al. [46] included an automated approach to derive precise railway alignment (centreline between two tracks) from LiDAR scans in order to create IFC alignment objects from that. The authors used LAS files (without colour) created by ground vehicle mounted LiDAR sensors (2000 points per square metre) to extract railway alignments and expressed them as IFC files (xBIM toolkit for IFC 4x1 was used). The point cloud was segmented in the direction of the trajectory using a distance filter in order to minimise the processing effort. An intensity-height-filter was applied to remove non-rail points. The extracted railway line was stored as polylines and the centreline between the rails was used to create the IfcAlignment instance to form an IFC file. The validation was carried out comparing the generated alignment using the reference points for the construction.

The reconstructing of 3D railway geometries from surveying data often depends on highly detailed point clouds that can even serve as a basis for the extraction of overhead cables. All of the presented works required highly detailed point clouds and have often created the data sets for the sole purpose of reconstructing highly detailed geometries. Some of the reviewed work has eventually converted the 3D reconstructions to the open IFC data model. None of the found work used GIS data to support the semantic segmentation process.

Hüthwohl et al. successfully demonstrated the applicability of mapping a 2D texture to a 2D plane for a beam face of a 3D IFC model [17] using an image as a 2D texture in 3D space, albeit in the context of classification of Building Information Modelling (BIM) of damages and exploring their suitability to include inspection and integrating damage information modelling for RC bridges in a standardised and open IFC format within an open bridge management system [47, 18]. A prototype software was developed for that purpose, allowed by the use of “Gygax construction IT research platform for 2D & 3D” [48], that relied on IFCEngine DLL, Helix Toolkit and SharpDX, to allow the shader to render the texture as a UV-mapped (i.e., S,T in IFC) triangulated mesh consisting of split rectangle into two triangles to a linked image.

The authors highlighted several disadvantages to this approach due to problems with estimating the correct location and orientation of the image plane in the IFC model, the distortions resulting from taking close-up photos and increase in data size by embedding the texture into the IFC model. This furthermore includes the risk of corrupting the IFC file during data exchange, when an external Uniform Resource Identifier (URI) of the image embedded in the IFC model without sharing the attached texture or when the texture image is moved to a different location.

Several pilot attempts for formalising compatible texture generation and integrating it within the entities of the IFC schema have been discussed in grey literature and shown potentials for BlenderBIM’s capability to handle texture maps [49]. However, no attempt has been found in the body of literature so far to generate UV-mapped textured meshes of irregular 3D shapes in IFC automatically, mainly due to the complex set of challenges to be faced from aliasing problems arising from inadequate sampling of all colour frequency values for texels through interpolation of texture attributes projected onto an intermediary arbitrary parameterised surface. Mipmapping based on Level of Development (LoD) requirements could be a solution in this case, yet it would require generating at least 3 texture maps of different resolutions based on the required triangle area of the triangulated faces defined for each LoD 3 and upwards.

• •

  Examining the potentials for cost savings by using freely available GIS data in early planning phases of railway projects by estimating horizontal and vertical alignments from the inferred semantic segmentation of the predicted railway trackbeds in GIS-compatible data format (i.e., GeoJSON) that could be later used as input for automatically generating and IfcAlignment or querying multi- or bi-temporal differences along a buffer zone of the alignment.

• •

  Examining the feasibility and compatibility of progressive meshing and surface mapping from vertex coloured point cloud to generate UV-mapped textured meshes of complex irregular 3D shapes in IFC automatically.

Usually, the LiDAR PCD consists of the x, y, z coordinates of points, CRS Projection Information and little to no semantic classification to them. Often, there are also no colours like RGB-values assigned to the points. Furthermore, PCD is often documented within tabular or textural formats, such as .csv, .xyz or within the widely used open format LAS, which may include some semantic classification, depending on the age, technical specifications of the scanning device and its support for the latest LAS format version. The targeted PCD converted from a LiDAR scan in this study usually comes from surveying flights and is therefore, unlike data from other kinds of acquisition, recorded from above and therefore involving specific limitations that need to be addressed, which will be discussed in section 4.

In the first step, the orthophotos are used to colour the PCD, in the case of missing RGB values of the respective points. Orthophotos result usually from surveying flights and are typically provided as georeferenced, grid-based image data, such as .(geo)tiff.

The GIS data, together with its classification of GIS features, can be derived from querying GIS databases, such as OpenStreetMap or provided GIS data from public institutions. It is important to identify the necessary label and later object classes in advance to decide on which object classes should be represented in the subsequent as-is model. The used GIS data consists usually of 2D polygons, sometimes containing attributed information about the height of the real object, for example in the case of buildings. The labelling data that comes with the 2D GIS features, such as the classification into “road”, “vegetation area”, “railways”, etc., is used in a next step to create annotation masks. Especially when using PCD originating from aerial scans, points with a higher z-value can cover other points of a different class, such as treetops covering a street when viewed from above. The GIS data helps to form coherently connected segmentation areas while also including algorithms of nearest neighbour to avoid false labelling. More information regarding this can be found in section 4.

The annotation masks can now be used in a 80:20 training vs. validation dataset relation within the Deep Learning (DL) model, namely 2DPASS [41]. The trained model can then be applied via inference to selected case study data, which is formerly unknown to the model. The outcome of applying the trained DL model to real data is an inferred segmentation from the already trained model based on a thresholded probability map. The resulting segmentation can be segregated into preassigned classes for reducing processing time and resources and their respective point clouds can then be treated individually.

Using the results from the previous step, the post-processing steps can be executed. This includes 3D mesh reconstruction from the class-specific point clouds with semantic information coming from the initial GIS data. Depending on the class, the result leads to 3D meshes of buildings in LoD2 or higher depending on the required Level of Information Need (LoIN) within a given coordinate reference system. The final meshes are furthermore first instanced as individual objects and then converted to a suitable data model, such as IFC and textured according to the orthographic photos.

The base dataset from LiDAR scans provided in the LAS data format. It consists mainly of three layers: ground, above ground and outliers (i.e., Ground, 20 and Unclassified in Thuringian LAS files). Table 2 details the default classes available in LAS format. As relying solely on the cadastre masks without further preprocessing steps was found insufficient to correctly annotate points related to specific classes like points related to external walls and corners of buildings, roads and railways that may lie outside the perimeter of the shape instances, which cannot catch their fine details. Therefore, to automate the annotation process of the other classes, each tile is processed with Connected Components Extraction and Density-based spatial clustering of applications with noise (DBSCAN) to cluster the point cloud and rely on the initial clustering for set operations with overlaid masks from the ATKIS relevant shapes that are geometrically manipulated to create buffer layers for annotation with the overlapping label extracted from their attributes’ table.

The connected components labelling with CloudCompare operates by propagating a front on the surface implicitly represented by the point cloud with the Fast Marching algorithm and making it dependent on scalar distance values associated to each point and neighbouring entities within an octree structure; which is followed by smoothing via Gaussian distance field gradient [50]. A sample of the dataset original PCD and orthophotos and the underlying processed clustering interim results for the semi-automated annotation are showcased in fig. 3.

The cadastral shape files follow a very consistent and detailed data schema that is managed by the Working Committee of the Surveying Authorities of the States of the Federal Republic of Germany (AdV) and disseminated through the Documentation on the Modelling of Geoinformation of Official Surveying and Mapping (GeoInfoDoc; German: Dokumentation zur Modellierung der Geoinformationen des amtlichen Vermessungswesens (GeoInfoDok)) [51]. The adopted and currently used version 7.1.2 was used to query the relevant attributes from the transport mode attributes for rail and road and thus facilitate the automatic annotation of the data set.

The data structure is designed to store and organise GIS data within the so-called AAA schema to facilitate managing and querying spatial data [51]. AAA stands for AFIS/ ALKIS/ATKIS; which refers to the official reference-point network, cadastre data and topographical maps. The geographical information contains the actual spatial data including geometric representations of geographic features such as points, lines, and polygons, which are stored in the Shapefile format (i.e., .shp). The associated attributes or properties of the geographic features are stored as key-value pairs with information about the keys and the values’ units derived from the GeoInfoDoc. The geodetic Coordinate Reference System (CRS) is ETRS89 for UTM zones 32N and 33N in the States of Thuringia and Saxony (i.e., EPSG 25832 and 25833) respectively. Information about the coordinate reference system used for the geometries, including its name, authority, and parameters. Since the number of attributes queried was very limited, there was no direct attempt to use prior knowledge about indexing the data to optimise for efficient spatial queries. For roadway axis object “Ax_Fahrbahnachse” the following attributes were queried as listed in table 3.

Whereas in railways the attributes of the railway shape objects “AX_Bahnstrecke” and “AX_Bahnverkehr” were queried for the following key-value pairs as listed in table 4. The derived attributes’ values from the ATKIS masks are then used for set operations on the clustered above ground LAS layer with an optimally derived empirical threshold of $>65\%$ overlapping of clusters to automate the labelling of class to balance precision and computing speeds to extract the vegetation, elevated roads, bridges and railway classes. Thresholds below this value would trigger a further step of sorting for the nearest label from cluster centre that annotates the largest sample of cluster points within a buffer zone of 0.75x the mean width in 2D space [52]. The following classes could be identified using the aforementioned approach:

1. 1.

   Vegetation: high and low.

2. 2.

   Road: traffic roads with their pavements, footpaths and parking lots.

3. 3.

   Railway: trackbeds including sleepers and ballast, with subclasses for rails, posts and cables could be extracted, but were merged together eventually in one class for balancing reasons.

4. 4.

   Crossings: bridges over land, water bodies or railway as well as tunnels and either assigned to road or railway class.

5. 5.

   Buildings.

6. 6.

   Miscellaneous class: “undefined” for found objects that fit none of the aforementioned categories, like sparse bird flock in Jena, vehicles in Erfurt, ships and boats in Dresden, transmission towers in Weimar, etc.

For the generation of 2D masks from orthophotos, two methods have been tested: (1) Using overlay set operations with derived geometries from the ATKIS shapes on synthetically generated connected components labelled images, which originate from the Gaussian mixture model fit of the pixels RGB values of the original orthophotos in colour space as features and (2) trials of instance segmentation via inference with zero-shot generalisation from the Segment Anything Model [53]. Especially in cases of pale vegetation and buildings with low contrasting coloured roof tops to the surroundings, (2) resulted in better image clusters for annotations.

The processed dataset included 313 grid cells of 1x1 kilometres from the cities of Erfurt, Jena and Weimar in Thuringia and 236 from the cities of Leipzig and Dresden of 2x2 kilometres in Saxony respectively with a total number of 1,257 cells after quad-splitting the grid cells from Saxony to 1x1 kilometres. Every cell was further quad split with the number of points in each block chosen at 4096 for training with limited GPU resources. The following classes were annotated: buildings, vegetation, water bodies, roads, railway trackbeds and a miscellaneous class for none of the aforementioned categories.

The custom dataset had to be manipulated and converted through an intermediary MMDetection3D format to provide description of the 3D boxes with 600 epochs for training of modified 2DPASS architecture. The best trained model achieved a semantic segmentation result of 71.48% for the Mean Intersection over Union (MIoU). Further sensitivity analysis steps to optimise the training hyper-parameters are needed and an increase in number of points per block to 8192 should be investigated to determine whether the segmentation result can yield a higher value for the evaluation metric.

This step takes the segmentation results along with the colourised PCD sampled from orthophotos to construct 3D primitive shapes and enrich them with generated context from the derived segmentation and related GIS data or meshes 3D surfaces with texture with a focus on IFC compatibility to be integrated in the information model.

The first use-case is demonstrated by generating an estimation of the segmented track-bed of the railway to roughly estimate its alignment. A 2.5D raster of 5000x5000 pixels resolution is generated to facilitate the estimation through classical image processing methods as shown in fig. 7. First, the image is thresholded to generate a binary image and cleaned from noise and thin elements with opening and closing. Second, the euclidean distance transform (EDT) is calculated to estimate the orientation and magnitude of the gradients of the EDT map.

Third, the gradient orientation is used as a prior guess of the vector normal to the tangent direction along the median axis. Fourth, the axis itself for the foreground is extracted using morphological skeletonisation and further processed with smoothing and cleaning of noise and small features below 2 metres (i.e., $\approx 5$ pixels). Fifth, the skeleton is split into separate labelled entities by detecting junction and endpoints, removing the junctions temporarily to disconnect the branches of the skeleton from each other to allow for connected component labelling (CCL) then reassemble them to their connected labels. Sixth, the line segments are extracted using the probabilistic Hough transform.

Seventh, circular segments of the skeleton are similarly estimated with a search interval [450, 10,000] with 5 metres step. The circular segments are extracted from the full circles with a boolean bitwise and operation on the original skeleton then estimated into polar coordinates through the endpoints of every segment to allow for ordered retrieval of each curve’s pixel indices from its branch in the minimum spanning tree. Eighth, the remaining skeleton parts are decimated through Visvalingam-Whyatt algorithm and fitted into clothoids with the $G^{1}$ Hermite interpolation based on the end points and the estimated conformed tangent vector normal to the gradient vector.

Finally, the parameterised line, circular and clothoidal segments are modelled as stringlines in a geodataframe and exported into GeoJSON format to be later used as an input to automatically generate the vertical and horizontal alignments in IFC using IfcOpenShell.

The second use-case is demonstrated on the Church of our Lady (Frauenkirche) found at the lower right corner of the cell in fig. 9 is used to demonstrate the possibility of integrating 3D coloured texture after meshing and automated UV-mapping of the vertices’ colours to a texture atlas into IFC using Xatlas. IfcOpenShell, BlenderBIM and FreeCAD are used to parse the mesh, texture atlas and the UV-map from the Wavefront OBJ format into their relevant IFC entities to generate a simple test project that contains the meshed instance as an IfcBuilding and the local path to the texture PNG image linked into the IFC file through a URI. Figure 9 shows the IFC model visualised in BlenderBIM alongside the embedded unwrapped UV-Map of the model shown in shown in the UV-editor window on the left. The UML Diagram elaborates the relations defined between the various entities of the textured building.

The difference between the recreated horizontal alignment ($s$) and the provided shape from ATKIS ($t$) using the Root Mean Square Deviation (RMSD) and the coefficient of variation (VC) to the RMSD are calculated using eqs. 1 and 2. Over a total length of 3870.444 metres, the RMSD and the VC thereof normalised by the mean value of the measured euclidean distances ($\delta\textsubscript{i}$) are calculated to be $12.506$ metres and $\pm 1.143$ respectively.

While the RMSD of fitting the recreated alignment in our presented workflow is higher than already published results in the literature [13, 15], it still has minimal impact on the precision of consequent output results when deployed for early planning applications. Furthermore, it provides a comprehensible pipeline to define nodes, edges, dependent features, labels and initialisation weights from the CSR Graph of the alignments, which lays the ground for designing a more refined regression solution based on a Graph Neural Network (GNN) framework to be addressed in a future work to derive recreated vertical and horizontal alignments optimally.

The demonstrated framework shows the automation of laborious manual as-is modelling, relying solely on openly available point cloud and GIS data. Especially in early or pre-planning phases of infrastructure projects, as-is models hold the potential to support decision-making processes without spending money on costly surveying tasks, leaving more time and budget for continuative thorough analyses. Combined with already established methods of quantitative difference estimation in bi-temporal scans [64], relevant queries about changes in context of railway can be collected and further processed to gain relatively accurate and up to date knowledge with minimal additional cost provided that the technical expertise is available. The industry standard IFC makes the results accessible both for further planning based on the created as-is models as well as allowing textured visualisation and documentation in an open format with a suitable data model. Public providers of cadastral and surveying data could, on the same data base they provide now, allocate detailed classified 3D models of the (built) environment to the public. With this, they would enable potential for many possible use cases in the context of infrastructure planning and at the same time relieve the public hand from avoidable expenses.