Year: 2024
Type: dissertation
Abstract: Rural fires present a significant danger to both human lives and the environment. This research aims to improve fire management and firefighting strategies by utilizing computer vision (CV) and deep learning to develop a close-range rural fire detection framework. The focus is on using low computational cost devices capable of localizing and classifying fires. The study also explores the impact of near-infrared (NIR) images on model performance and compares the effectiveness of two key techniques: Convolutional Neural Networks (CNN) and Transformers. The proposed framework comprises two object detector models, YoloV8 and EfficientDet, followed by a classification module. Two CNN-based models and three Transformer-based models were used to evaluate this module. The input data for the detectors consists of a newly created dataset containing "Fire" and "No Fire" images, compiled from five other datasets. The outputs from the detector models are then used as inputs for the classification module. The results indicate that the MobileNet, EfficientNet, and MaxViT models exhibited the highest performance. Specifically, mobilenet_v3_small_pt and efficientnet_v2_s_pt achieved a 99% F1 score, while maxvit_t_pt reached a 98% F1 score, showcasing their strong suitability for this task. In conclusion, the developed framework demonstrated robust performance, with NIR images significantly enhancing model accuracy. The findings suggest that smaller networks are adequate for this specific task, making the framework suitable for deployment on low-computation resource devices. This adaptability enables the use of the framework, including the two detectors and one classifier, in microprocessors and robotic systems, considering the overall floating-point operations per second required.
Author: Dias, Filipe de Correia Pedro Marçal
Language: English
Subfield: Safety, Risk, Reliability and Quality
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7043047616