Research Of The Video Fire Detection Algorithm Based On Spatio-temporal Domain Texture

This thesis proposes a video fire detection system based on spatial and temporal texture features. We combine image processing, pattern recognition, computer vision, machine learning and other technologies to analysis the video image’s contents and target behaviors. Finally the automatic fire detection and alarm are accomplished.This thesis analyzes the motion features and color spaces to obtain the candidate fire pixels in the video image. Contrast experiments show that using interframe difference method under RGB and HSI color spaces for choosing candidate pixels performs better. In order to get more candidate fire pixels, we use seed-growth to extend the candidate pixels around the previous ones. Since flame is a fluid movement, we divide the video image into blocks, then select candidate fire blocks according to the number of candidate fire pixels in each blocks.Several existing texture feature extraction methods are analyzed in the video fire detection. Based on feature fusion technology, we combine the covariance matrix with wavelet transform and Contourlet transform respectively to get the spatio-temporal texture feature methods, and apply them in the video fire detection system. Four different kernel functions are analyzed in support vector machine (SVM) learning and classification of the texture feature. The experimental results show that the classification based on the linear kernel function is better than others. In order to determine whether the image block for inspection is fire or not, mapping function is employed to classify the feature block which is extracted from the video image. According to the fire flame evaluation index, whether the current image has fire is determined at last.Finally, this thesis concludes with experimental results based on LBP, LBP-TOP, covariance matrix, covariance matrix and discrete wavelet transform fusion, covariance matrix and Contourlet transform fusion, which indicate that the proposed method outperforms other fire detection algorithms, providing high reliability and a low false alarm rate.