<![CDATA[Fire is a destructive event that causes significant losses; therefore, an accurate and rapid early fire detection system is essential. With advancements in computer vision technology, deep learning-based object detection methods are increasingly employed for real-time fire detection. This study aims to evaluate and compare the performance of You Only Look Once (YOLO) based models specifically YOLOv11, YOLOv12, and YOLOv26 in detecting fire and smoke objects in real-time using cameras, and to identify the optimal model based on evaluation results. The dataset consists of fire imagery categorized into two primary classes: fire and smoke. All three models were trained using identical configurations: an image size of 640 × 640 pixels, a batch size of 8, and 100 epochs on Google Colab. Performance was evaluated using precision, recall, F1-score, accuracy, mAP, and confusion matrices. Training results indicate that YOLOv11 achieved the highest mAP@50 value (approximately 0.52), compared to YOLOv12 (0.47) and YOLOv26 (0.50). In real-time testing, YOLOv11 achieved 93.3% accuracy, followed by YOLOv26 at 92.6% and YOLOv12 at 90.11%. Based on these results, YOLOv11 demonstrated the best performance under the experimental conditions, achieving the highest real-time accuracy and mAP@50 for both object classes. However, smoke detection remains a major challenge for Convolutional Neural Network-based fire detection systems due to smoke's dynamic, semi-transparent, and low-contrast characteristics. Consequently, the YOLOv11 model holds potential as a foundation for developing effective real-time fire detection systems, such as those for embedded systems, intelligent surveillance, and UAV-based platforms.]]>
