Fire Detection Using YOLOv8: A Simple Study with YOLOv11 Comparison

Abstract

Fire accidents are still very common, and in many cases the damage becomes worse mainly because the fire is not noticed at the right time. Most traditional alarm systems react only after smoke or heat reaches the sensor, and by then the situation may already be risky. Because of this gap, I tried to make a simple fire-detection setup using YOLOv8 that can directly look at images and recognise fire almost instantly. For this work, I used a small collection of fire and non-fire images which had different backgrounds and lighting so that the model does not depend on only one type of situation. The idea was not only to see how YOLOv8 performs but also to understand how it stands compared to the newer YOLOv11 model, which many recent studies mention. I did not train YOLOv11 in this project, but I referred to its improvements just to get an idea of where YOLOv8 is strong or weak. From the tests I did, YOLOv8 responded quickly and did not raise many false alarms in normal conditions. It could mark the flame as soon as it appeared in the frame. Overall, this small study shows that YOLOv8 can be used for basic fire detection in simple setups. It is fast, light, and easy to run, and can be improved further with more data or by trying newer versions like YOLOv11 later. This make the approach useful for places where a quick visual warning system is needed.

Introduction

Fire accidents remain common and often become destructive before anyone notices, largely because traditional fire-alarm systems depend on heat or smoke reaching a sensor. This delay in detection—especially in crowded buildings, industrial zones, and areas with heavy electrical wiring—allows fires to spread rapidly. In countries like India, many fire incidents escalate due to a lack of early monitoring.

To address this, modern approaches increasingly use vision-based fire detection through CCTV cameras. Computer vision can analyze video frames in real time and identify early signs of fire or smoke faster than conventional sensors. Early rule-based methods relied mainly on flame color or motion but were unreliable, often triggering false alarms due to reflections, sunlight, or dynamic scenes.

The introduction of deep learning, particularly convolutional neural networks (CNNs), improved accuracy by learning flame features directly from images. However, CNNs were too slow for real-time monitoring and could only classify images, not locate fire regions. YOLO-based object detection models addressed these issues by detecting fire and providing bounding boxes in a single fast pass. Versions such as YOLOv3–YOLOv8 have been widely adopted in smart cities, factories, construction sites, and wildfire monitoring due to their speed and accuracy.

YOLOv8, with its anchor-free design and efficient feature extraction, offers strong real-time performance and better detection of irregular flame shapes. Newer versions like YOLOv11 show improvements in stability and small-object detection. Vision-based systems also benefit from low hardware cost, since they can run on standard CCTV cameras.

The extended work evaluates YOLOv8 and compares it conceptually with emerging architectures like YOLOv11 to understand their performance in early fire detection, focusing on response time, false alarms, and detection reliability under normal lighting.

The literature review shows the evolution of fire-detection approaches—starting from color and motion-based image processing, progressing to CNN-based models, and ultimately moving toward high-speed YOLO architectures. Studies consistently highlight YOLO models as the most effective for real-time applications.

In the methodology, traditional CNN models are described as limited because they classify frames but cannot localize fire. YOLO solves this issue by detecting objects in a single step. YOLOv8 is the primary model used, and its architecture consists of a CSP backbone, PAN-FPN neck, and anchor-free detection head. These components allow it to detect flames of various sizes, shapes, and flickering patterns. YOLOv8 offers high accuracy, speed, and compatibility with low-power hardware, though it still struggles in smoky, low-light, or reflective environments.

Conclusion

In this work, I mainly tried to see how a simple fire detection system can be made using YOLOv8, without any complicated setup. I just used a normal webcam and some basic frames, and surprisingly the model reacted quite well. Whenever the candle flame appeared, YOLOv8 picked it up almost immediately, and when there was no fire, it didn’t show anything. This part was actually nice because false alarms are very irritating in real-life situations, and at least in my small test, it didn’t behave in a strange way. While doing the tests, I noticed that YOLOv8 works fine when the lighting is normal. But in slightly darker areas or when the flame was too small, the model hesitated a bit or the confidence score went down. So it’s not perfect, but for a basic trial, it was okay. When I checked the confusion-matrix part, the results more or less matched what I saw: good detections in simple scenes and a few misses when the situation was tricky. YOLOv11 was not used here, but I still mentioned it because many papers say it performs better with small objects and in difficult conditions. Maybe if this project is continued later, YOLOv11 can be tried to see if the small-flame problem gets solved or not. Right now, I only discussed it to show how newer versions might improve things. Overall, the study basically shows that YOLO models can be used to make a simple fire-detection idea work, especially when someone wants something fast and not too expensive. If more data, more lighting variations, or even outdoor tests are added later, the system can be improved. So this project is more like a starting point, and there is a lot that can be done in the future to make it more stable and reliable.

References

[1] Kumar P. M., Fire Detection using Deep Convolution Neural Networks in Video Streams, 2022. [2] S. Kamalanathan, Smart Surveillance System for Abnormal Activity Detection using CNN, 2021. [3] H. Qian, A Fire Monitoring and Alarm System Based on Channel-Wise Pruned YOLOv3, 2022. [4] A. Robert Singh, Fire Detection by Parallel Classification of Fire and Smoke Using Convolutional Neural Networks, 2021. [4] Y. Zhang, P. Geng, and C. Sivaparthipan, “Big data and artificial intelligence-based early risk warning system of fire hazard for smart cities,” 2021. [5] L. Zhao, Fire-YOLO: A Small Target Object Detection Method for Fire Inspection, 2022. [6] S. S. Muhammad, Fire Detection Using DenseNet-201 Algorithm and Surveillance Camera Images, 2024. [7] S. Ke et al., Improved YOLOX Fire Scenario Detection Method, 2022. [8] D. C. Prabhu T., “Automated Home Security using Facial Recognition with IoT Remote Monitoring,” IEEE IPACT, 2021. [9] L. Lou et al., “Smoke root detection from video sequences based on multi-feature fusion,” 2022. [10] S. Wang, ES-YOLO: A New Lightweight Fire Detection Model, 2023. [11] C. Li, Research on Fire Detection Algorithm Based on Deep Learning, 2022. [12] A. M. O. et al., “Model of settlement evacuation based on imitation modeling application,” IOP Conf. Ser., 2020. [13] F. M. Talaat, An Improved Fire Detection Approach Based on YOLO-v8 for Smart Cities, 2023. [15] S. H. Park, Recognition of IoT-based Fire Detection System Fire-Signal Patterns Applying Fuzzy Logic, 2023. [14] S. Bashambu, Real-Time Fire and Smoke Detection System, 2023. [15] S. G. Zhang, Swin-YOLOv5: Research and Application of Fire and Smoke Detection Algorithm, 2023. [16] D. Zhang, A YOLO-Based Approach for Fire and Smoke Detection in IoT Surveillance Systems, 2024. [17] A. S. Mahdi, Wildfire Detection System Using YOLOv5 Deep Learning Model, 2023. [18] S. Msedd, Fire Detection and Segmentation Using YOLOv5 and U-Net, 2021. [19] H. Wu, Ship Fire Detection Based on an Improved YOLO Algorithm with Lightweight CNN Model, 2022. [20] H. Bayer and A. Aziz, Object Detection of Fire Safety Equipment Using YOLOv5 Neural Network, 2022. [21] S. Yu, Forest Fire Detection Algorithm Based on Improved YOLOv5, 2022. [22] S. Kumar, YOLOv4 Algorithm for Real-Time Detection of Fire and PPE at Construction Sites, 2021. [23] A. Bochkovskiy, C-Y. Wang, and H-Y. Liao, “YOLOv4: Optimal speed and accuracy of object detection,” arXiv:2004.10934, 2020. [24] Y. Kumar, H. Gupta et al., A Novel YOLOv3 Algorithm-Based Deep Learning Approach for Waste Segregation, 2020.

Copyright

Copyright © 2025 Namrata Kaushik. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.