Options
SmartFire vision: Advancing fire detection in smart cities by hybrid deep Learning technique
Date Issued
2025
Citation
Iqbal, A., Azhar, M., Shafqat Ali, M., Usman, M., Wattoo, W. A., Farhan, M. (2025 Aug 26). SmartFire vision: Advancing fire detection in smart cities by hybrid deep learning technique. AI-SI 2025 - IEEE International Conference on Artificial Intelligence for Sustainable Innovation: Shaping the Future with Intelligent Solutions, Seri Pacific Hotel Kuala Lumpur, Kuala Lumpur.
Type
Conference Paper
Abstract
Recently, the risk of fire has globally been increased due to climate change, urbanization and tremendous growth of the population. Conventional smoke and fire detectors face
challenges in detecting and providing useful information about the location, spreading speed, and the size of a fire. With the boom of Al-based techniques, deep learning models
have shown promising results in addressing such challenges. This study presents a novel approach for smoke and fire detection in video sequences, utilizing an integrated Vision
Transformer (ViT) and DEtection TRansformer (DETR) model. Our approach leverages the Removing Inefficient Attention Heads technique within VIT to enhance feature extraction
efficiency and accuracy. Concurrently, DETR provides precise object localization and context understanding. The combined features of ViT and DETR are fused and classified
using a fully connected neural network, incorporating a thresholding mechanism to reduce false alarms and an integrated alarm system for rapid response. The proposed model
was trained on the Furg fire dataset and the results have validated the performance of our algorithm by achieving an overallaccuracy of 91.37%, recall of 85.27%, precision of
88.55%, and an F1-score of 86.64%.
Availability at HKSYU Library
