감시 카메라를 이용한 딥러닝 기반의 화재 경보 시스템

Abstract

Fires can cause significant human and property losses in unexpected locations. To detect fires, sensors based fire detection systems are being popular. However, the sensors should be carefully maintained because the sensor based detectors can malfunction due to humidity and dust, resulting in high cost. Recently, using surveillance cameras installed many places, vision technologies are being used to detect fires. However, the methods are highly dependent on empirical parameters, which cannot be designed systematically. As another approaches, frame differences are utilized to detect flames. However, these methods may have low performances in the environment of the low communication speed of the camera. This paper proposes new fire detection systems based on convolutional neural network. Considering limited number of dataset and the deformation of the camera images, datasets are augmented using rotations, vertical/horizontal flips, brightness deformation, image enlargement and reduction, etc. The proposed neural network uses the minimal size of the kernel, i.e. 3x3 because receptive fields should be minimal to detect small fires. Performances of the designed neural network are carefully evaluated using cross validation methods, such as confusion matrix based recall/precision/accuracy/f1-score, receiver operating characteristic (ROC), and area under the curve (AUC). In the image test, it is verified that the proposed neural network detects fires well in the normal environment including exceptional cases, i.e. irregular fire shapes, flaming sunsets, and small fires occupying 2% of the hole image. The achieved performances are f1-score=0.962 and AUC=0.98.