Computer Vision Approaches for Extracting Fire-Safety Information from Service Drawings

Read full text Summarise with VIRAT
Authors: Julius Ekstrand, Victor Truong
Cited by: 0
References: 23

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Fire-safety documentation is an essential part of regular inspections during the building operations and maintenance phase. As help, service drawings exist to detail the positions of various fire-safety devices and their addresses, as well as fire zone coverage throughout the floors and sectors of a building. Automatically extracting this information would be immensely helpful towards, for example, maintaining up-to-date documentation or integration onto digital systems. To that end, this thesis adapts multiple computer vision techniques in order to extract pertinent information: (1) a Keypoint R-CNN model is trained on a custom dataset to detect the symbols of fire devices along with their installation positions, (2) OCR is used to parse addresses and (3) a region merging procedure is implemented to segment fire zones. Additionally, symbols are also assigned to their addresses using the Hungarian algorithm, with Euclidean distance as base cost. For the symbol detection, we find that crucial model settings had to be changed in order to tailor the model for the domain and achieve satisfactory performance. High accuracy is achieved despite major class-imbalances in the dataset, but the model fails in all cases to detect the true installation positions of devices, whenever these are indicated by an auxiliary line. Assigning symbols and addresses to each other based on distance works well in most cases, while having a few edge cases that are impossible to amend with our method alone, due to inherent quirks in the service drawings. Using region merging for the fire zone segmentation gives us high recall, but low precision. The low precision is mainly due to the oversegmentation step of region merging, and partly due to the existence of hard-to-ignore false positives in most service drawings. Overall, the results across all tasks are promising while still leaving plenty of room for improvement, or alternative approaches, in further work.

Fields of Study & Subjects

Handwritten Text Recognition Techniques Advanced Neural Network Applications Image and Object Detection Techniques

Publication details

Journal / source
Lund University Publications Student Papers (Lund University)
Publication date
1 January 2026
Language
EN
Document type
Thesis