Video-Based Fire Detection with Saliency Detection and Convolutional Neural Networks

Much work has been done in fire detection by using color model and hand-designed features. However, these methods are difficult to meet the needs of various fire detection scenarios. In this paper we propose a new method of video-based fire detection by combining image saliency detection and convolutional neural networks. Our method consists of two modules: (1) utilize saliency detection method to extract flame candidate region proposals. (2) extract features from each candidate region by using convolutional neural networks, and then classify these features into fire or non-fire. This method can automatically learn effective features from video sequences. The experimental results show that our method achieves classification results superior to some hand-designed features for fire detection. We also compare color model method and saliency detection method for obtaining flame candidate regions.