Search

Fire detection has attracted increasing attention from the public because of the huge loss caused by fires every year. Compared with the traditional fire detection techniques based on smoke or heat sensors, the frameworks using machine learning methods in videos for fire detection have the advantages of higher efficiency and accuracy of detection, robustness to various environments, and lower cost of the systems. The uniqueness of these frameworks stems from the developed machine learning approaches for autonomous information extraction and fire detection in sequential video frames. A framework for flame detection is proposed based on the synergy of the Horn-Schunck optical flow estimation method, a probabilistic saliency analysis approach and a temporal wavelet analysis scheme. The estimated optical flows, together with the saliency analysis method, work effectively in selecting moving regions by well describing the dynamic property of flames, which contributes to accurate detection of flames. Additionally, the temporal wavelet transform based analysis increases the robustness of the framework and provides reliable results by discarding non-flame pixels according to their temporally changing patterns. Apart from the dynamic characteristic of flames, the property of colours is also of crucial importance in describing flames. However, the colours of flames usually vary significantly with different illumination or burning material, which results in a wide diversity. To well model the various colours, a novel flame colour model is proposed based on the Dirichlet process Gaussian mixture model. The distribution of flame colours is represented by a Gaussian mixture model, of which the number of mixture components is learned from the training data autonomously by setting a Dirichlet process as the prior. Compared with those methods which set the number of mixture components empirically, the developed model can access a more accurate estimation of the distribution of flame colours. The inference is successfully implemented by two methods, i.e., the Gibbs sampling and variational inference algorithms, to manage different quantities of training data. The colour model can be incorporated into the framework of flame detection and the results show that the colour model achieves a highly accurate estimation of the distribution of flame colours, which contributes to the good performance of the whole framework. All the proposed approaches are tested on real videos of various environments and proved to be capable of accurate flame detection.