A Real-Time Approach to Fire and Smoke Detect...

A Real-Time Approach to Fire and Smoke Detection with YOLOv5 and Streamlit

Author Name : G. Nagappa, Mehtar Jeelan, K. Govardhan Reddy, G. Harshavardhan Sai, Kb. Venu Gopal

ABSTRACT Efficient smoke detection is crucial for minimizing fire-related damages, yet existing technologies often suffer from low detection rates and high false negatives. This paper presents a unique smoke detection model based on an optimized version of YOLOv5. By assembling a diverse dataset comprising both real and synthetic smoke images, the model is trained and fine-tuned using different YOLOv5 variants (s, m, l) and loss functions (GIoU, DIoU, CIoU). To address the challenge of limited training samples, a mosaic enhancement technique is employed to generate augmented images. Additionally, a dynamic anchor box mechanism is introduced to adaptively adjust anchor box sizes and positions during training, overcoming inaccuracies inherent in YOLOv5. Furthermore, attention mechanisms are incorporated to improve feature map balance across different scales. Comparative analysis demonstrates a notable 4.4% increase in mean Average Precision (mAP) over the baseline model, with a detection speed of up to 85 frames per second (FPS), making it highly suitable for practical deployment.