Multiple-Scale Fire Detection Based on One-Stage Deep Learning Model

Automatic fire detection study is very important to prevent the risk of fire. Recently, most of fire detection systems are included in unmanned surveillance system instead of using additional temperature sensors. Recent development in deep learning such as convolutional neural network(CNN) significantly improved the fire detection performance in the sense of both accuracy and speed. However, existing CNN-based fire detection methods still fails to detect a small-scale fire. In this paper, we propose a multi-scale fire detection model based on Yolo one-stage detector which shows an acceptable performance in speed and accuracy. Moreover, we train the model using both existing and in-house fire dataset that includes various indoor environment such as room and hallway. Experimental results show that our model works well within a fully automatic surveillance system to detect a fire accident with a higher accuracy in real time.