Early and reliable detection of smoke and fire is essential for minimizing damage and ensuring public safety in smart city environments; however, existing vision-based approaches often suffer from limited contextual understanding, high false-alarm rates, and reduced robustness under complex visual conditions. Conventional convolutional neural network and single-stage object detection models primarily rely on local features, which restrict their ability to detect early-stage or visually diffuse smoke patterns. To address these limitations, this study proposes an intelligent vision-based fire and smoke detection framework that integrates a Vision Transformer (ViT) with the YOLOv8 detection architecture. In the proposed model, ViT is employed as a global feature extractor to capture long-range spatial dependencies and contextual relationships, while YOLOv8 serves as a real-time detection head for precise localization and classification of fire and smoke regions. Experimental evaluation is conducted on the Fire and Smoke Dataset and the Forest Fire Smoke Dataset, comprising over 7000 diverse images collected from urban and rural environments. The proposed approach achieves an accuracy of 99.2%, precision of 98.5%, recall of 97.8%, and an F1-score of 98.1%, representing an improvement of approximately 4.3% in accuracy compared to conventional CNN-based and YOLO-only methods. Additionally, low inference latency demonstrates the suitability of the model for real-time deployment. These results confirm that the proposed ViT-YOLOv8 framework effectively overcomes the limitations of existing approaches and provides a robust, accurate, and scalable solution for early smoke and fire detection in smart city applications.
Abozeid et al. (Sat,) studied this question.