Research On Smoke Detection Algorithms For Large Space Video

With the development of science and technology,surveillance video has been widely deployed in various large space scenarios.Monitoring video can not only realize real-time viewing,recording,playback,recall and storage of video images,but also its emerging applications hava been explored by researchers at home and abroad,such as fire detection and pedestrian detection and so on.These methods can not meet the needs of fire detection in large space scenarios because of their low accuracy,instability and technical limitations.Video fire detection(VFD)technology has better stability than traditional contact fire detection technology.It can be used in warehouse,station,square and other space scenarios.Therefore,it has been widely concerned.This thesis mainly studies some key technologies of VFD in large space scenarios.On the basis of existing VFD technology,uniform local binary pattern(ULBP),discrete wavelet transform(DWT),machine learning classifier and convolutional neural network(CNN)theory are introduced into VFD and further studied.The following work is mainly completed for the study:1.A smoke motion region detection algorithmth based on YUV color space smoke filtering rules and gauss mixture model(GMM)motion detection is proposed by studying some commonly used moving object detection methods.These methods can not effectively extract the complete smoke moving area.And the extracted area contains a large number of non-smoke pixel objects in complex environment.Because the smoke area can not be extracted effectively,the computation of image classification algorithm will increase directly,which will affect the real-time performance,and even more will seriously lead to inaccurate image classification and false alarm.The method proposed by the thesis effectively captures the motion area and filters out non-smoke pixels in the area.2.A smoke image recognition method that based on multi-feature fusion and real adaboost is proposed by studying some common feature extraction methods and machine learning feature classifiers.The features obtained by using only one feature extraction method are very unstable and vulnerable to environmental interference.Moreover,smoke information in smoke images can not be extracted adequately by using only one method.In addition,in most of the existing smoke recognition methods,only the commonly used machine learning classifiers for feature classification are used.Therefore,the classifier algorithms of these methods may not match the input features,which result low recognition accuracy of the final smoke image algorithm.The method proposed by the thesis combines multiple features,and the matching algorithm with higher matching can effectively improve the recognition accuracy of smoke images.3.A smoke image recognition network model based on convolution neural network is proposed by studiyng some classical convolution neural network model.In the existing image recognition methods of VFD technology,the shallow machine learning mode is used for smoke image recognition.It uses feature extraction method to obtain smoke information in the smoke image to form feature vectors,and then uses machine learning classifier feature vectors to distinguish,so as to achieve the purpose of smoke image recognition.In shallow machine learning,the features in images are selected manually,which is insufficient or redundant.The network model proposed by the thesis has a simple network structure and can automate image feature extraction and image classification.