An Improved Fire and Smoke Detection Method Based on YOLOv7

Fires can cause extensive damage to human life and property, highlighting the importance of accurate flame and smoke detection systems. However, current methods for detecting flame and smoke struggle to balance the demands of real-time processing and prediction accuracy, limiting their applicability in fire warning and detecting systems. In this paper, we propose a novel approach based on the YOLOv7 architecture for efficient and accurate detection of fire and smoke in images. Our approach incorporates partial convolutional layers into the E-ELAN module of the YOLOv7 network, enabling faster and more precise identification of fire and smoke. Furthermore, we introduce the Focal-EIoU loss function to address the issue of fluctuating model loss caused by low-quality samples. We validate our approach on a real-world dataset and report significant improvements in detection accuracy, with a mean average precision of 78.5 and an FPS increase to 63. These results demonstrate the effectiveness of our approach in enhancing the capability of fire and smoke detection systems.

I. Introduction

Fire poses a significant threat to people's lives and property safety. According to data from the Ministry of Emergency Management of China, there were 703,000 cases of fire incidents, resulting in 1,557 deaths, 1,769 injuries, and direct property losses of 5.85 billion yuan from January to Octo-ber 2022 alone. However, manual firefighting by firefighters remains the predominant method of combating fires, which often leads to firefighter fatalities. Therefore, the development of firefighting robots is essential for enhancing firefighting efficiency and safety. Firefighting robots primarily rely on camera-based detection systems to identify fire and smoke, making rapid and accurate fire and smoke detection in video images crucial.