Fire Detection and Segmentation Using YOLOv8-Seg Based on Computer Vision

Abstract

Fire is a disaster that can cause enormous material and life losses, while conventional smoke and heat sensors respond slowly and have limited detection range. This study develops a fire detection and segmentation system based on the YOLOv8n-seg architecture, an instance segmentation model capable of producing precise pixel-level fire masks so that small/static fire objects (e.g., candles, matches, stoves) can be distinguished from large/dynamic fires (e.g., building or forest fires) to reduce false alarms. The model was trained using transfer learning on a manually annotated dataset of 240 training images and 60 validation images, with aggressive geometric and HSV color augmentation to address the limited dataset size. Training was performed for 50 epochs on a Google Colab NVIDIA Tesla T4 GPU. Evaluation results show a mask segmentation precision of 98.0%, recall of 96.4%, and mAP@50 of 98.2%, with per-class AP@50 of 99.50% for small/static fire and 96.92% for large/dynamic fire. The exported model (6.45 MB) achieved a real-time inference speed of approximately 30.4 ms per frame (~33 FPS) when tested on a live webcam stream, indicating its feasibility for real-time fire early-warning applications.