Performance Benchmarking of YOLOv11, YOLOv12, and YOLOv26 for Real-Time Fire Detection

Abstract

Fire is a devastating event that causes significant losses. Therefore, an early fire detection system that can work accurately and quickly is urgently needed. As computer vision technology advances, deep learning-based object detection methods are increasingly being used for real-time fire detection. The purpose of this study is to test and compare the performance of You Only Look Once (YOLO)-based models, namely YOLOv11, YOLOv12, and YOLOv26, in detecting fire and smoke objects in real-time using cameras. The data used are fire images with two main classes, namely fire and smoke. All models were trained using the same parameters, namely an image size of 640 × 640 pixels, a batch size of 8, and 100 epochs on the Google Colab platform to ensure an objective evaluation process. Performance evaluation uses precision, recall, F1-score, accuracy, mean Average Precision, and error matrix. The training results of YOLOv11 have the highest mAP@50 value of around 0.52 compared to YOLOv12 at 0.47 and YOLOv26 at 0.50. In real-time testing, YOLOv11 achieved the highest accuracy of 93.3%, followed by YOLOv26 at 92.6%, and YOLOv12 at 90.11%. YOLOv11 showed the best balance between precision and recall in both object classes. However, smoke objects remain a major challenge for Convolutional Neural Network-based fire detection systems due to their dynamic, semi-transparent, and low-contrast nature compared to their surrounding environment. This study demonstrates that the YOLO model allows for effective implementation in real-time fire detection systems and has potential for further development in embedded systems, intelligent surveillance, and UAV-based platforms.