A Study of Novel Initial Fire Detection Algorithm Based on Deep Learning Method

Abstract

A small ember, created by a chemical reaction between a substance and oxygen, can grow into a large fire as temperature, wind, and weather conditions change. A growing fire incident can have devastating consequences, including property loss, environmental damage, and loss of life, which is why early fire detection is so important. There are various detection devices such as smoke detectors, heat detectors, fire detectors, and gas detectors that can be used in the early stages of a fire. While early fire detection system developments incorporating IoT technology are emerging in various industries, Smoke alarms, the most common type of smoke detector in homes and offices, accounted for 96.6% of all malfunctions from 2021 to July of the previous year, totaling 249,445 incidents. The analysis of detector malfunctions showed that non-fire alarm factors such as food, cooking, and dust accounted for the largest share of 40.6%. This paper proposes an algorithm for early fire detection by incorporating a deep learning-based model to compensate for the problem of non-fire warning malfunctions, which is a shortcoming of existing detectors. Finally, for fire detection, a bounding box for the fire is specified using a smoke detector, a thermal imaging camera, and a webcam camera trained with the Yolov7 model. Then, we propose an algorithm to remove the bounding box of non-fire reports and malfunctions from the heating map using smoke detectors and thermal imaging cameras. After applying the algorithm proposed in this paper, only fires with heat sources are recognized, and all bounding boxes for non-fire reports are removed.