Learning multi-layer ICA features for video-based fire detection

Previous work on fire detection has focused on hand-designed features or carefully designed detectors. However, there are no universal hand-designed features or detectors that work well for various classification tasks or even for various fire detection scenarios. In this paper we propose a new method of video-based fire detection by learning multi-layer ICA spatiotemporal features. This method can automatically learn effective features from 3D spatiotemporal blocks with combination of ICA learning algorithm, stacking and pooling. A linear SVM is used in classification stage. The experimental results show that, despite its simplicity, our method achieves classification results superior to some popular spatiotemporal features for fire detection. We also compare different block sampling strategies and different ways of feature construction for our method.