The Study Of Fire Smoke Detection Based On Video Image Sequence

Fire is a serious threat and hazard in the world.It is a very important issue to be able to warn fire in a timely and accurate manner.There are two main types of fire detection,which are based on flame or smoke detection.The thesis mainly studies smoke detection technology.Traditional sensor-based smoke detection technology has many limitations,which prompts people to study smoke detection algorithm based on video surveillance,and puts forward many smoke detection algorithms.Most of the smoke detection algorithms achieved good detection performance,b ut there are still two main problems.In the middle and late stages of fire,the detection effect of the algorithm is very good.However,in the initial stage of fire where the smoke is thin and the smoke’s moving speed is slow,the performance of the algorithm is not satisfactory.In addition,most of them were based on hand-crafted features or used raw images directly as input to the neural network,which led to poor robustness of the algorithm and did not have universal applicability for different smoke scenarios.To solve the above two problems,two kinds of smoke detection algorithms based on video image are proposed in the thesis.1.In order to improve the detection performance of the initial fire,a novel smoke detection method based on local extrema co-occurrence pattern and energy analysis is proposed in the thesis for the scene of thin smoke and slow movement in the initial fire stage.The method mainly consists of three phases,namely preprocessing stage,feature extraction stage and image classification stage.In the preprocessing stage,the motion foreground area of the video frame is extracted by using the algorithm,and the smoke pixels are identified by the color space using in the motion foreground area.Then,the texture features are computed using local extrema co-occurrence pattern and the smoke energy feature vector is computed using smoke energy statistics,the two features are normalized and fused to a feature vector.Finally,the fused vector trains the support vector machine to recognize the smoke.The experimental results show that the method can detect the smoke generated in the early stage of the fire in time and effectively,and reduce the loss caused by the fire.2.In order to improve the robustness of the smoke detection algorithm,a smoke detection method based on convolutional neural network is proposed.Firstly,adopt the normalization method to remove influence of illumination,and the suspected smoke regions are detected using smoke color model.Then,the features of suspected region is extracted automatically by convolutional neural network,on that the smoke identification is performed.Finally,the corresponding alarm signal is obtained by the identification results.For the problem that the region of smoke is relatively small in the early stage of smoke generation,the strategy of implicit enlarging the suspected regions is used to improve the timeliness of smoke detection.To solve the overfitting caused by insufficient training samples or imbalanced,the data enhancement techniques are used to generate more training samples from original training data sets.The experimental results indicate that the method can effectively detect smoke in a variety of complex scenarios,so it has higher accuracy and better robustness.