RUB » Bibliotheksportal

Jan-Hendrik Heinbach, Angelina Aziz

• This paper addresses the limitations of manual fire safety inspections in buildings, which are typically conducted by fire safety managers using inspection checklists. These inspections are usually time-consuming and prone to errors. To mitigate these issues, a proof-of-concept implementation is proposed to partially automate the inspection process. Various visual inspection aspects were identified, such as completeness, accessibility, and visibility of fire safety equipment, based on legal requirements. This object detection approach uses state-of-the-art neural models (YOLOv7) to visually inspect documented images with respect to the identified inspection aspects. The models are trained on self-created datasets to detect and inspect fire extinguishers, emergency exits, fire and smoke protection doors, and call points. By adopting the proposed method, several benefits can be realized, including increased automation, reduced time requirements, and the potential for facility owners to conduct inspections themselves. Furthermore, it is possible to expand the implementation to encompass more visual inspection aspects and targets. Additionally, integrating different data sources, such as BIM (Building Information Modeling), can significantly broaden the range of inspections in future studies.