BUILDING A FIRE ALARM DEVICE USING YOLO AND SENSORS

Abstract

In the first six months of 2024, there were approximately 70.600 fires reported worldwide and around 2.222 fires in Vietnam. These alarming statistics emphasize the growing risk and underline the urgent need for enhanced fire prevention and safety measures. Therefore, we believe that a fire alarm device is essential. In this thesis, we have developed a device that utilizes computer vision and machine learning, combined with sensors, to detect fire and issue warnings. The device is designed to function as a warning camera installed in the kitchen, providing continuous monitoring to detect potential fire hazards in this high-risk area. The dataset was built with three classes: fire, smoke and safe-fire. We trained and evaluated multiple models, including YOLOv8n, SSD MobileNetV2, and SSD EfficientDetD0, to determine the most suitable model for this project. The results indicated that YOLOv8n outperformed the other models, making it the best fit. Further training achieved a mAP50 of 73%. The YOLOv8 model was converted to the NCNN format to reduce the inference time from 582ms of YOLOv8n to 117.3ms of NCNN. It was then deployed on the Raspberry Pi 4 Model B, integrated with both a temperature sensor and a smoke sensor, enabling real-time fire detection and warning functionality.