Generating ‘as-is’ 3D models of existing infrastructure There has been a lack of viable ‘as-is’ 3D modelling methods for existing infrastructure. Researchers have now recently fine-tuned the first scene and object detection method; capable of automatically modelling infrastructure from arbitrary 3D points while also estimating the layout and connections of the objects in 3D. Industrial Technologies © Eva Agapaki A major challenge faced by engineers today is that of effectively mapping and labelling existing infrastructure. Currently, over two thirds of the effort is taken up with manually converting clouds of points (coordinates) to a 3D model. The end result is few constructed facilities with a complete record of as-built information, information which is not produced for the vast majority of new construction and retrofit projects. It is estimated that this has led to rework and design changes costing up to 10 % of the original installation cost. The EU-funded INFRASTRUCTUREMODELS project was set up to improve the situation by developing a means to detect and classify common building objects from visual and spatial data, reducing the time needed to create the as-built geometric Building Information Model (BIM). The project successfully created methods that generate slab, wall, room, and floor objects/assemblies in point cloud datasets, quickly generating as-is 3D models of existing buildings, reducing costs. Exploiting the ‘architectural priors’ As-built models are virtual, 3D object-oriented models of existing structures. Their difference with an as-designed model is that their dimensions are a true reflection of what is actually built. Traditionally, attempts to render ‘as-is’ 3D infrastructure models have employed a somewhat bottom-up approach as they locate significant points, then find significant lines and planes, before moving on to objects. This however does not work so well for many infrastructural elements. As the project’s coordinator Dr Ioannis Brilakis, explains, “Many infrastructure objects are ‘boring’. Columns, beams, slabs and walls for instance have very few significant points to distinguish them from each other.” To get around this challenge the INFRASTRUCTUREMODELS project focused on the contextual information that can make such objects distinguishable; their position in space, alongside their relationship to other objects. They achieved this feat by creating a set of segmented stages that broke down a building’s point clouds into floors, rooms, walls/slabs, etc. As Dr Brilakis summarises, “Buildings are complex, but they all have floors, walls, doors and other key elements that have relatively well defined and standardised relationships with each other. These architectural priors are what we are trying to exploit to reduce complexity and assist the detection process.” The team used image analysis tools to visually characterise the infrastructure, identifying and numerically mapping the elements. These representations, along with their inferred topological relationships, constitute the ‘elements’ used by machine learning algorithms to create meaningful infrastructure categories for the models. The models can then automatically detect element types, with their relative positions, from a range of viewpoints. By estimating their distance to the observer and their 3D position in space relative to other points, these elements can be mapped onto the 3D point clouds and rendered with colour and texture to generate accurate models. As Dr Brilakis further enlightens, “Images are very useful in determining what category each object belongs to. For instance, a white colour on a vertically-aligned long and thin cuboid increases our confidence that we are looking at a wall object. A carpet pattern on that same object would infer something different.” The main challenge faced by the team was in collecting enough data to train the algorithms, with infrastructure often ranking low amongst all Big Data sectors. With points clouds being no exception, they found very few industry partners had them available to share. Furthermore, out of those available, most did not meet the required quality standard. The project team is confident that their 3D models will benefit building tasks such as retrofitting energy infrastructure to fix problems, increase efficiency or improve green credentials. Another prime example of likely applications is for building operations management. Additionally, the impact from shut down management of industrial plants could be minimised, if as-built models of the plants were available. The team is currently in the process of setting up a commercialisation centre in Cambridge, UK, intended to boost the rapid development of this technology, alongside other associated advances. The aim is to bring products to market within two-three years. Keywords INFRASTRUCTUREMODELS, infrastructure, as-is models, 3D, retrofitting, engineering, cloud points, architecture, image analysis, algorithms, operations management
