Development of a smart edge device for fire detection

• Main Author: Ng, Wei Yuan
• Format: Final Year Project / Dissertation / Thesis
• Published: 2023
• Subjects: TJ Mechanical engineering and machinery
• Online Access: http://eprints.utar.edu.my/5828/1/MH_1801391_Final_NG_WEI_YUAN.pdf; http://eprints.utar.edu.my/5828/

Conventional fire detection systems based on smoke and flame sensors have been shown to be unreliable in terms of accuracy, response time, and false alarms. This project proposes a smart edge fire detection system that overcomes the limitations of conventional fire warning systems by utilizing deep learning models and edge computing. The system is based on an object detection model for fire detection using the Improved YOLOv5s algorithm, which integrates BiFPN and an additional prediction layer for detecting small targets of fire. The system is designed to be deployed on edge devices, specifically the Jetson Nano B01, and includes a user-friendly interface accessible via Telegram. The workflow of the project is divided into three sections: Google Colab, a workstation, and the Jetson Nano B01. The Google Colab section is used to train the models and generate the Improved YOLOv5s model, while the workstation is used to download the trained model weights and generate the Flask server and MQTT client. The Jetson Nano B01 section is used to install the dependencies, set up the system, optimize the trained model with TensorRT and TorchScript, and connect the Flask server to NodeRED. The final system is capable of starting with user input and sending alerts to Telegram when the fire is detected, with end users able to access the system via Telegram to receive alerts and view the video stream. The results show that the Improved YOLOv5s model is the best option, with an average FPS of 10.5 and a high recall of 0.5, indicating a high capture rate of fire, although with lower precision of 0.541. The latency of the edge computing system is 554ms, making it an efficient and reliable option for detecting fires in real time. Future works could explore the use of a larger dataset and ensemble models to increase accuracy, optimize the model for devices with constrained resources, and improve the user interface and messaging service integration. Overall, this project contributes to the development of efficient and reliable fire detection systems that can be deployed in various settings.