Video Smoke Recognition Based On Optical Flow Characteristics And Wavelet Transform

Fires have caused great damage to human beings,how to correctly identify the fire has important practical significance.The traditional smoke detection method using sensor is strongly influenced by the environment,however,video smoke detection method doesn’t require much hardware and has better practicability.Conventional identification methods often choose more stable wind situation,however,under strong wind conditions,the flame burning fast,smoke alse shows more features.This paper considers the characteristics of smoke under this conditions,thus achieve better early-warning effect.The first operation is to filter and adjust video image pixels,then compare the advantages and disadvantages of three extraction methods of movement through the experiment,Gaussian mixture model is chosen as the final foregroud extraction method.Then using histogram statistics to get smoke pixels’ distribution in RGB and HSI color space,thus using color features to extract smoke pixels,then the scattered blocks are turned into connected regions after dilation and erosion operation,and the small areas are removed after the connected domain operation.Finally using smoke boxes mark suspected area.For smoke under the condition of strong winds,its status is not stable,so it is need to analyze the corresponding characteristics of the smoke.Optical flow method can get the basic motion vector information of smoke pixels,several characteristics of smoke are obtained by analyzing the distribution of these values.Due to the influence of wind,the movement trend between the majority of the pixels and the entire region in video image is consistent,so orientation consistency feature is gotten.By analyzing the distribution and fluctuation information of direction and speed of pixels,the mean and variance characteristics of length and direction of smoke flow are acquired.To classify the speed of pixels within smoke region and compare them,the feature of the contrast of optical flow is extracted.Finally,introduce the transparent characteristics of smoke to get the wavelet energy feature.After comparison and analysis the corresponding features between smoke and none smoke,the feature vector set for classification and detection is collected.In order to reflect the best classification results of the features,this paper analyzes the performance of three classifiers.K nearest neighbor classifier needs to calculate the distance between a sample and all the other samples,the computation is verylarge.Neural network convergence is slow,with the presence of complex nonlinear system,its predictive ability is poor.The support vector machine have the obvious advantages of fast processing speed and strong ability of nonlinear data processing.Finally libsvm toolbox is used as classifer.By building classifier model with training set,any video images is detected.Finally,in order to verify the effectiveness of the proposed algorithm,the algorithm in this paper is compared with previous classic recognition algorithm,under the conditions of the same training set and test set,the comparison results of both accuracy and error rate are obtained,the experimental results show the method can detect wildfire smoke effectively and has better strong anti-interference ability,robustness and wide application.