YOLOFlame: Benchmarking vehicle fire and smoke detection in urban environments

Fire incidents are highly destructive events that cause loss of life, environmental and economic disruption, property damage and even contribute to climate change. Detecting fires or smoke at an early stage is crucial for mitigating these losses and minimizing their associated impacts. Vision sensors have emerged as a promising research area for fire and smoke detection, attracting the attention of experts in the field of computer vision (CV). Initially, fire and smoke detection relied on low-level color features; however, more effective deep learning (DL) models have replaced these, achieving higher accuracy. These models still face challenges, such as a high false alarm rate (FAR), because they consider fire and smoke detection as a categorization task, disregarding the localization of fire and smoke in a given scene. Additionally, their time complexity and model size delay real-world implementations. To address this challenge, we introduce a lightweight detection approach, YOLOFlame, built on the YOLOv8 nano-architecture integrated with an attention module. The proposed method achieves competitive performance while significantly reducing model complexity and size. It effectively detects both small and large fire and smoke regions in the images. Furthermore, we present a manually annotated medium-scale dataset titled as Inferno, which includes six distinct classes. The proposed YOLOFlame achieved an mAP of 0.66 on the newly curated Inferno dataset, while having a smaller number of trainable parameters and faster training time. • Proposed a novel lightweight YOLOFlame method for fire and smoke detection. • Proposed a novel manually annotated dataset for fire and smoke detection entitled as Inferno dataset. • Extensive quantitative and qualitative evaluation of the proposed YOLOFlame method.