University of Twente Student Theses


Deep learning on 3D point clouds for safety-related asset management in buildings

Anjanappa, Geethanjali (2022) Deep learning on 3D point clouds for safety-related asset management in buildings.

[img] PDF
Abstract:Buildings are equipped with multiple safety-related assets depending on the need and functionality. It is vital to have up-to-date information on these assets for emergencies. As a part of infrastructure maintenance, asset management systems document assets within a building, maintain their records, and continuously monitor them to improve the asset's performance. In this regard, an asset management system can establish a centralized system for safety-related assets in buildings to (i) enable finding them quickly and effortlessly; (ii) keep up-to-date records for first responders; (iii) monitor the entire safety-related infrastructure; (iv) keep a check on the building's compliance with safety standards. A crucial step in establishing such a system for safety-related assets is to identify these assets within the building. In collaboration with CGI Inc., this research explored the scope of using 3D point clouds and the Deep Learning scene segmentation approach to identify safety-related assets within buildings. We adapt the Kernel Point-Fully Convolutional Network (KP–FCNN) to perform scene segmentation to identify safety-related assets. The research focuses on common assets in most buildings like ceiling lights, exit signs, ventilation ducts, windows, doors, stairways, fire switches, and extinguishers. We use point cloud datasets acquired from three different 3D sensors to evaluate the designed method, namely, the depth camera (S3DIS), a Mobile Laser Scanner (HPS dataset), and a consumer-grade lidar sensor (iPhone dataset). In addition to standard evaluation metrics, asset identification rate (AIR) is used to evaluate the rate of correctly identified asset instances. The results from various experiments showed that our workflow could successfully identify small-sized assets like fire switches and exit signs with a 100% AIR in some cases. The designed method also proves invariant and robust by successfully performing scene segmentation for the three chosen datasets to identify safety-related assets. Additionally, based on qualitative analysis, we determine that it is possible to identify safety-related assets within a building using the point clouds obtained from simpler lidar devices, like iPhone.
Item Type:Essay (Master)
Faculty:ITC: Faculty of Geo-information Science and Earth Observation
Subject:38 earth sciences, 54 computer science
Programme:Geoinformation Science and Earth Observation MSc (75014)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page