کنفرانس مهندسی برق ایران

صفحه اصلی / سی و چهارمین کنفرانس بین المللی مهندسی برق

Advancing Robotic Fire Suppression: A Multi-Source, Fine-Grained Visual Fire Detection Framework with Hard Negative Mining

نویسندگان :

Amirmahdi Froughinia¹ Sadra Rafatnia² Elahe Sadat Abdolkarimi³

1- Sahand University of Technology 2- Sahand University of Technology 3- Tafresh University

کلمات کلیدی :

Robotic Fire Suppression،Fire Detection،YOLO12m،Hard Negatives،Real-time Inference

چکیده :

Visual fire detection in robotic fire suppression systems requires a model that, in addition to flame detection, is capable of fine-grained resolution of potential ignition sources and discrimination of fire-like visual patterns. To address this need, a comprehensive framework for multi-source and fine-grained fire detection is presented, which is designed based on a large-scale data-set with 43,825 images in eight specialized classes and about 4,000 structured negative samples. Negative samples are collected by focusing on deceptive patterns including warm lights, orange reflections, and LED glows to ensure the robustness of the model in real environments and reduce false positives. Three architectures YOLOv5s, YOLO11m, and YOLO12m are evaluated under identical and controlled training conditions, and the results show that YOLO12m provides the best performance, achieving mAP@0.5 of 0.982 and mAP@0.5:0.95 of 0.895. Analysis of the learning behavior and decomposition of the cost function components—including Distribution Focal Loss, IoU-based regression loss, and classification loss—show that the model exhibits a more stable decision boundary and higher discriminability than fire-like patterns in the presence of hard negative examples. The proposed model is then deployed in an end-to-end operational chain including conversion to ONNX and TorchScript formats, GPU-based real-time inference, and a robotic alerting subsystem. The findings indicate that the proposed approach elevates fire detection from a reactive process to a practical framework for early fire hazard prediction and provides direct applicability in robotic fire-fighting systems.