Smart Early Fire and Smoke Detection with Transfer Learning

In this study, a low latent edge-based fire and smoke detection and alert system was designed to detect and prevent indoor fires caused by human or machine faults, or lack of appropriate fire safety measures. To live comfortably in this fast-paced society, safety is a must. To detect early signs of fire, smoke and flammable gases, a real-time fire detection system was developed by combining a Raspberry Pi and a Raspberry Pi camera v2. Images taken from an open-source fire and smoke dataset were annotated and fed to the models. The model selection involved training YOLOv4-tiny, YOLOv5s TensorFlow lite pre-trained models, and deploying the most feasible transfer learned model having a higher mean average precision (mAP) on Raspberry Pi, which acts as the edge device. It performs real-time analysis of live feed captured by a Raspberry Pi camera, sends email alerts to the user stating confirmation of the presence of fire or smoke, and prevents false alarms caused due to the data collected by IoT devices by conducting a double check with the captured feed for detection. The user can act upon this notification and ensure timely preventive action. The proposed state-of-the-art indoor alert system achieves a mean average precision (mAP) of 42.3 for fire, 35.4 for smoke classes; a precision of 58.5 for fire, 64.3 for smoke and a recall of 45.9 for fire and 33.3 for smoke respectively. If the system is improved in the future, it is expected that it will be able to give faster and more accurate alerts.