Occlusion-aware object detection under fire and smoke conditions for intelligent surveillance

Traditional object detection mechanisms in surveillance systems tend to fail under occlusion-dense scenarios due to fire, smoke, or environmental noise and thus achieve poor reliability in emergency response scenarios. In this paper, the authors introduce a multi-stage deep learning framework that incorporates fire-aware preprocessing and occlusion-resistant object detection with YOLO-based architectures. The pipeline starts with a Fire Detection Module that uses a custom-trained YOLO variant on an annotated fire dataset of 2000 images to detect active fire and smoke areas in video streams. Merged and padded bounding boxes around these areas are expanded by 60% to account for visibility loss because of occlusion, an empirically derived hyperparameter that has been optimized using ablation analysis. The extended regions are fed into a YOLO segmentation model learned from a 3291-image annotated dataset of 21 vehicle components to support fine-grained occluded object detection. Experimental performance indicates dramatic improvement in mean Average Precision (mAP) and F1 scores over typical YOLOv8 baselines, especially for fire-obscured vehicles. Confusion matrix evaluation and F1- confidence curves confirm the model's performance in sustaining classification accuracy at low confidence levels (F1 ≈ 0.84 at 0.36 confidence). This fire-optimized, occlusion-conscious detection architecture significantly enhances the reliability of intelligent surveillance systems operating in low-visibility, high-threat contexts, and it is applicable to deployment in urban traffic surveillance, industrial security, and smart city infrastructure.