Automated BIM-based modeling of lighting elements in buildings with computer vision methods
DATE:
2019-07-22
UNIVERSAL IDENTIFIER: http://hdl.handle.net/11093/1306
UNESCO SUBJECT: 2209.08 Iluminación
DOCUMENT TYPE: doctoralThesis
ABSTRACT
Lighting is a key factor in the design and maintenance of a building, both in terms of costs and comfort. Approximately 19% of all the electricity consumed globally corresponds to lighting, with artificial lighting accounting for one-third of the total electricity used in buildings. Moreover, this consumption has been recently increasing at a rate of 2.4% per year; this manifests a clear need for a more efficient use of the lighting resources.
Applying effective energy-saving strategies to improve the energy efficiency of the lighting installation of a building requires a complete and accurate determination of its lighting inventory and conditions. But modelling and simulation in this domain have been marred by the lack of accurate information. Therefore, a precise and automatic system enabling the aforementioned inventory process becomes crucial to reduce energy consumption in buildings. Several methods can be used to implement this automatic system; among them, computer vision is one of the most commonly used technologies to solve automatic recognition problems. Computer vision systems (CVSs) can be used to realize this process, but some requirements have to be taken into account to provide an adequate solution, namely cost, speed, detection performance and reliability.
The collected data from the automatic system has to be included in the digital representation of the building, and building information modelling (BIM) is one of the most widespread frameworks to store and manage this information in the architecture, engineering and construction (AEC) industry. BIM is a valid tool to manage the digital representation of all the aspects of the building, integrating design and project information of its entire lifecycle. BIM can be used solely to include the lighting information, but it can also be used to extract useful information about the building that can be leveraged by the automatic detection system and improve its performance.
This thesis presents novel research work targeting a comprehensive automatic system for the detection, identification and localization of lighting elements in buildings. For this purpose, the thesis proposes the use of computer vision methods to automatically perform the detection process based on localized grayscale images and information extracted from the BIM model of the building. The generated information is automatically fed back and integrated into the BIM model, using it as both input and output for the methods.
The proposed system is initially envisioned as a three-step process: (i) image and geometric analysis, (ii) clustering and averaging, and (iii) insertion in BIM. The first iteration of the system features techniques to perform an initial candidate search on the images and highly reduce the effective processing area, improving the speed of the methods in the rest of this first step. Then, in the second step, a clustering is performed and lighting objects are positioned by averaging both the location and orientation of the individual elements in each cluster. Finally, this information is used to enrich the BIM model of the building in the last step. Subsequent modifications are later introduced in each of the main steps to better use the available BIM data, resulting in increased performance and enhanced versatility and reliability of the detection system.
These improvements are empirically evaluated in several case studies with a total of five different lamp models and more than 30,000 test images. The results obtained in the experiments on the initial candidate search techniques yield a reduction of up to an order of magnitude in processing time and memory usage. Moreover, the final version of the system is able to detect more than 12,000 lamp objects within the source material, achieving an accuracy of 99.93% of correct identifications and an average localization error of 13.63 cm. These figures evidence the suitability of the method for the intended use cases, and the extended applicability and increased performance achieved thanks to the proposed modifications.