Enhancing Real-Time Fire Detection with RT-DETR and Optimized Dataset Preparation

Myla, Joseph Raju (2024) Enhancing Real-Time Fire Detection with RT-DETR and Optimized Dataset Preparation. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (500kB) | Preview
Preview	PDF (Configuration Manual) Download (1MB) | Preview

Abstract

Real-time fire detection systems face significant challenges achieving high accuracy and efficient processing speeds. This research looks at the optimization of dataset preparation techniques that could be used to implement RT-DETR in fire detection systems, together with augmentation strategies. The presented study aims to bridge the critical gap between dataset preparation methodologies and the performance of transformer-based architectures in safety-critical applications.

In the work, a holistic approach was followed with the RT-DETR-L architecture. Extensive data augmentation is done through geometrical transformations, including changes in intensity. The implementation is performed on Google Colab, running on an A100 40GB GPU infrastructure. It comes with a dataset of 2,200 validation images. The pipeline is carefully designed in such ways that the aspect ratios of the images are preserved, standardized at 640x640 resolution.

Very remarkable performance metrics are presented, with 0.985 for mAP@50 and 0.949 for mAP@50-95. The system also retains very high precision regarding fire detection at 0.991 and smoke detection at 0.962, with only 16.6ms of processing time per image. This represents substantial outperformance compared to the current benchmarks while preserving real-time processing capabilities.

These results provided new baselines for real-time fire detection systems and came with useful insights into the optimization of dataset preparation for transformer-based architecture. The contribution of this study both in theoretical understanding and practical implementation strategies advances the development of enhanced fire detection systems, which shall be very operational, especially in application areas concerning the safety of human life where a real-time response is expected.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Rifai, Hicham UNSPECIFIED
Subjects:	Q Science > QA Mathematics > Electronic computers. Computer science T Technology > T Technology (General) > Information Technology > Electronic computers. Computer science Q Science > QH Natural history > QH301 Biology > Methods of research. Technique. Experimental biology > Data processing. Bioinformatics > Artificial intelligence > Computer vision Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Artificial intelligence > Computer vision
Divisions:	School of Computing > Master of Science in Data Analytics
Depositing User:	Ciara O'Brien
Date Deposited:	03 Sep 2025 14:46
Last Modified:	03 Sep 2025 14:46
URI:	https://norma.ncirl.ie/id/eprint/8755

Actions (login required)

View Item