Study And Realization Of Fire Detection Based On Video Image

The fire will cause a huge loss of human finances,destroy the ecological environment where human beings live,and even directly threaten human life.For these reasons,the detection of fire has become the focus of attention.Its appearance and development have very important significance for human development.Traditional methods of fire detection such as smoke detection and temperature detection are mainly based on sensor equipment.However,such detection methods are not applicable to buildings with large spaces,and there are many cases of false alarms.Fire detection based on video images make up for the disadvantages of traditional fire detection methods and are widely used in forest,subway,and large space fire detection.The application of algorithms such as morphology,image enhancement and image segmentation in video fire detection is discussed in depth in this thesis.The main research results of this paper are as follows:Aiming at the interference of moving objects such as vehicles and people,fire prediction based on Gaussian Mixture Model(GMM)and RGB color model is proposed.By fire prediction based on GMM and RGB,most of the interfering objects can be excluded.Fire prediction increases the utilization rate of resources and achieves the real-time detection effect.In the region growth image segmentation algorithm,all the pixels in the image need to be traversed to select the seed,and the seed selection is a waste of time.In this thesis,the GMM is used to select the suspected fire pixels,and the growth seeds are selected through the YCr Cb color model.The improved region growing segmentation algorithm in this thesis not only saves seed screening time,but also has better segmentation effect.According to the image features of the fire in the early stage,the features and extraction algorithms of fire are studied.On the basis of image preprocessing and image segmentation,the area,rounded degree and texture features of fire and their extraction algorithms are studied and realized.Then these features are used as the criterion of fire detection,and the fire detection and recognition work is completed.According to the features of the extracted fire,combined with support vector machine(SVM),the fire detection model is built.Because the parameters' selection of SVM is the key factor of SVM performance,this thesis optimizes the parameters of SVM with the fruit fly optimazation algorithm(FOA)that is simple and easy to implement.For FOA,it is easy to fall into local optimum defects.In order to overcome the defects such as local optimum in FOA,the fixed-step FOA is improved to a dynamic step-size and two-subgroup optimization algorithm called IMFOA,which enhanced the overall optimization capability and local search ability of the FOA.The IMFOA avoid premature convergence of algorithm and fall into a local optimum.The IMFOA improves the optimization ability and convergence speed of the original FOA.When the SVM model is optimized,the parameters precision of improved fruit fly optimization-based SVM model is improved.The experiments shows that the improved fruit fly optimization-based SVM can improve the accuracy of fire detection.