Fire Detection from Images Based on Single Shot MultiBox Detector

Fire detection is quite necessary because fire is very harmful to both lives and properties of humans. However, the causes of fire are various and complicated, which leads to the difficulty of fire prevention. As the increase of video surveillance systems, fire detection from images has become a research hotspot. Traditional algorithms focused on the contour, color and movement features of fire, which are only efficient in particular scenes and cannot be generalized. With the development of deep learning, convolution neural networks with good generalization performance have been broadly used in object detection. In this paper, a one-stage object detector based on deep neural network, namely Single Shot MultiBox Detector (SSD), is introduced to fire detection. In order to improve the generalization performance of detector and tackle the training difficulty of the lack of training samples, a dataset with a large number of flame and fire images, namely HHFire are constructed. The format of annotation is the same as that of Pascal Visual Object Classes. Images from the dataset covers lots of real scenes, including forest, buildings, fields, indoors, etc. Experimental results have demonstrated the efficiency of SSD on HHFire.