Open Collections

Abstract

The urgent need for advanced fire detection systems stems from the increased intensity in fire events that cause massive property losses and irreversible damage. To overcome the limitations of traditional fire detection systems such as smoke detectors, computer vision (CV) algorithms have been adopted to improve fire detection accuracy. A CV-based fire detection system takes images from fire monitoring camera as input and output fire-related information. To identify the installation location of fire monitoring cameras, this research performed compartment-level fire risk analysis based on fire risk analysis method for engineering (FRAME). Relevant fire risk factors were quantified using a building information modeling (BIM) model, and their relative importance was evaluated through the analytic hierarchy process (AHP). The resulting compartment-specific fire risk values provide data-driven guidance for deploying cameras in high-risk areas of a building. The case study result of a residential building aligns with historical fire incident statistics, which validates the proposed framework. In terms of fire detection model, when compared to single-modal model, multi-modal model has gained attention because it leverages the richer information presented in both RGB and thermal images. However, prevalent multi-modal fire detection methods significantly increase model complexity by requiring two separate streams in the backbone to process RGB and thermal images independently. To address this issue, this thesis proposes a 4-channel single-stream fire detection method based on YOLOv5, which concatenates RGB and thermal images to form the required 4-channel input. Comparison experiments with dual-stream YOLOv5 models using add fusion and transformer fusion demonstrate that the four-channel single-stream model reduces the model complexity while improving detection accuracy. To further enhance the accuracy of multi-modal fire detection and reduce model complexity, this study redesigned the YOLOv5's C3 module by integrating the Convolutional Block Attention Module (CBAM) to form the C3CBAM module and introduced the SCYLLA-IoU (SIoU) loss function. By comparing its performance with that of state-of-the-art models in multi-modal object detection such as YOLOv5-based dual-stream model, this study showcases that the proposed approach improves fire detection in diverse conditions presented in the selected dataset.