The Research On Fire Detection Based On Video Image

Traditional fire detection systems always have poor performance in response speed andreliability. With the rapid development of image processing and pattern recognition, firedetection based on video images has become one of the hottest research points recently.On the base of the analysis about the related researches, the basical algorithmsframework is proposed. And this paper delves into the research about the fire video imagepre-processing, possible fire region (PFM) segmentation, the flame feature extraction, andfeatures fusion.Firstly, some pre-processing methods for the flame image，such as image graying,Gaussian smoothing and median filter, are analyzed. Comparison shows that image graying isa suitable pre-process method for the subject here.Secondly, some methods for possible fire region (PFM) segmentation, such as edgedetection, thresholding, motion detection and color segmentation, are studied. Consideringthe problems encountered in the experiments, an improved temperature-area thresholdingalgorithm is proposed. Comparing with other methods, thresholding and color segmentationare more suitable for the fire detection. Experiments show that, the threshold selected by theproposed temperature-area thresholding algorithm is close to the optimal one; this algorithmnot only can select threshold adaptively to a certain extent, but also is valid for the flamerecognition.Thirdly, after the study on the static and dynamic characteristics of fire, this paperextracts the corresponding flame features in image: high gray value, color, circularity andflame cusp, contour shape change (dct_dis) and area change. The high gray value reflects theflame’s temperature feature and luminance feature; Roundness and the flame cusp depict theirregular characteristic of the flame contour’s shape; dct_dis indicates the flame shape’schange, and indirectly reflects the flame’s flicker.What’s more, the Bayes classifier and support vector machine (SVM) are studied.Simulations using the data extracted from the experimental video show that, the Bayesclassifier performs not very well in small sample classification; The C-SVM classifier withradial basis kernel function is a good choice for flame features fusion, but the choice of the RBF kernel and the punishment factor needs to be tested by more samples.Finally, the algorithm is implemented in the Fedora OS using C++. The validity of thecircularity feature and the proposed temperature-area thresholding algorithm are furtherdiscussed. The final results of the flame identification show that, the algorithm withoutclassifier has yielded a good performance, with lower missing rate and false positive rate.