A modified YOLOv5 architecture for efficient fire detection in smart cities

Abstract

Fire disasters are considered to be among the most harmful hazards, causing fatalities, ecological and economic chaos, property damage, and they can even impact climate change. Early fire detection is necessary to overcome these losses and disruption. Fire detection using vision sensors is a promising research area that has gained significant attention from computer vision experts. Traditionally, low-level colour features were used for fire detection but they have now been superseded by effective deep learning models that achieve higher accuracy. However, these models also suffer from a higher false alarm rate, due to the fact that they treat fire detection as a classification task where the entire image is classified into a single class and the region of the proposal stage is ignored. Furthermore, the time complexity and model size limit these models from real-world implementation. To overcome these challenges, we propose a modified YOLOv5s model that integrates a Stem module in the backbone, replaces larger kernels with smaller ones in the SPP (Neck) and adds the P6 module into the head. This model achieves promising results with lower complexity and smaller model size, and is able to detect both small and large fire regions in images. Moreover, we contribute a medium-scale fire dataset that consists of three classes (i.e. vehicle fire, building fire, and indoor electric fire), with manual annotation according to the object detection model. This dataset will be made publicly available for research purposes. Finally, for fair evaluation, we re-implement 12 different state-of-the-art object detection models, including the proposed model, and trained them over a self-created dataset. We found that in comparison the proposed model had better detection performance and applicable in real-world scenario. Our codes and dataset will be publicly available at https://github.com/Hikmat-Yar/Modified-YOLOv5-Code.