심층학습 영상분석 기반 화재 검출 시스템 구현

Abstract

The performance of convolutional deep learning networks is generally determined according to parameters of target dataset, structure of network, convolution kernel, activation function, and optimization algorithm. In this paper, a proper deep learning model and parameters for video-based fire detection were selected through simulations and applied the learning results to the fire detection solution. We compare and analyze the fire detection performance of AlexNet, GoogLeNet, and VGG-16 to select an effective network for detecting flame and smoke. The learning characteristics and the accuracy for flame and smoke dataset are analyzed according to the sizes and strides of convolution kernel. Dataset for training deep learning models is classified into normal, smoke and flame. Normal class images includes images with clouds and foggy. The kernel size is larger and the smaller the stride in kernel characteristics, the higher accuracy for the image dataset for fire detection. In terms of deep learning network structure, the accuracy of VGG-16 is better than that of other networks. We implement a fire detection solution based on Caffe framework that classifies flames and smoke frame from normal frame for the input video. As experiments of fire detection, it is shown that developed solution can be applied fire detection based on video.