A Study on Fire Detection Using Deep Learning and Image Filtering Based on Characteristics of Flame and Smoke

When a fire breaks out, damage to human health is more often caused by poisoning and suffocation related to the occurrence of smoke than by a direct cause such as exposure to flame. In addition, fire that is in the condition of smoldering has fatal potential for the human body because it shows a high rate of production of carbon monoxide rather than carbon dioxide due to an incomplete combustion process. Therefore, this study sought to achieve early image-based detection not only of flames but also of smoke in the event of a fire. To this end, a flame area was pre-processed using color and corner detection, while smoke could be detected using dark channel prior characteristics and optical flow. For the pre-processed region of interest, a deep learning-based convolutional neural network was used to infer whether the region was a fire. Through this approach, it was possible to improve accuracy by lowering the error detection rate compared to when a fire was detected through an object detection model without separate pre-processing. To evaluate the performance of the proposed method, inference was conducted through a directly photographed image. As a result, the an accuracy level of 97.0% in the case of flames and 94.0% in the case of smoke could be confirmed.