Abstract ▼ AI-Helper

In the development of new smart cities, there is a shift from conventional fire-detection systems towards vision-based systems, leveraging emerging technologies like digital cameras, computer vision, artificial intelligence, and deep learning. This thesis work focuses on the creation of a highly eff...

In the development of new smart cities, there is a shift from conventional fire-detection systems towards vision-based systems, leveraging emerging technologies like digital cameras, computer vision, artificial intelligence, and deep learning. This thesis work focuses on the creation of a highly efficient fire detection system capable of swiftly identifying even minor sparks and triggering an alarm within just 8 seconds of a fire outbreak. A novel convolutional neural network has been developed for the task of fire detection, employing an enhanced version of the You Only Look Once (YOLO) network. Building upon the improved YOLOv4/5/7 algorithms, this network has been adapted to function on the Banana Pi M3 board, utilizing only three layers. The initial experiments involved an assessment of the conventional YOLOv approaches to evaluate their accuracy in predicting candidate fire regions. However, these experiments did not yield the desired results in terms of fire accident detection. To address this, enhancements were made to the traditional YOLO networks. This included expanding the training dataset size through data augmentation techniques to facilitate real-time monitoring of fire incidents. The proposed method also involved modifications to the network structure through techniques like automatic color augmentation and parameter reduction. These modifications collectively resulted in a system that could effectively and rapidly detect and classify catastrophic fires in varying weather conditions, including sunny and cloudy days and nighttime scenarios. The experimental findings affirm that the proposed method holds significant promise for enhancing the safety of smart cities and monitoring fire incidents within urban areas. In conclusion, a comprehensive performance evaluation was conducted by comparing the effectiveness of this method with other recently reported fire- detection approaches, utilizing widely accepted performance metrics to assess the fire classification results achieved.