Studies On Dynamic Texture And Convolutional Neural Networks Based Smoke Detection In Video Sequences

Once a fire occurs,it will spread rapidly,devouring human life and wealth,and destroying the natural environment.Fire detection is important in fire prevention and it can provide great help for reducing disaster losses.Traditional temperature sensors and smoke sensors are widely used in fire detection systems,with the country's economic growth and development,people make higher requirements for fire detection.On the one hand,there are more and more places need fire detection,such as high-rise buildings like commercial complexes and terminal buildings,and outdoor areas like vast forests and pastures.On the other.hand,it is necessary to improve the detection effect,such as achieving earlier detection,visualization,etc.Video fire detection technology(VFD)is a promising method to solve these problems,especially in the context that deep learning drives breakthroughs in computer vision and artificial intelligence.Moreover,video smoke detection technology is a more efficient approach for early fire detection and expects to make great progress for practical engineering applications.This paper aims to deeply study the theory and methods of video smoke detection,and provide theoretical and technical support for the practical application of video smoke detection technology.At present,the research on video smoke detection method based on deep learning has just started.So it is still in the state of coexistence of traditional methods and deep learning methods.This paper has studied both methods.The specific research contents are as follows:(1)A dynamic texture smoke detection method based on irregular motion regions is proposed.Burning experiments using cotton rope and n-heptane as fuel are conducted in the standard room.Then the high-definition videos containing black smoke,white smoke and pedestrian interference are taken and form the smoke dynamic texture train-ing and test datasets.Firstly,sliding window are used for smoke detection.Motion area is extracted using background subtraction method.According to the motion area,suspected smoke blocks with size of 100 pixels x 100 pixels are generated.LBPTOP descriptors are used for dynamic texture feature extraction in suspected smoke blocks.Then SVM is used for training and classification.The test results show that the size of the blocks,motion area rate and frame rule have great influence on the performance.Therefore,we propose a dynamic textured smoke detection method based on irregular moving regions.The experimental results show that this method greatly reduced the false alarm rate while maintaining a high detection rate.In addition,we also compared the performance of the descriptors such as LBPTOP,VLBP and CVLBP in terms of coding mode and sample point number etc.(2)Propose a method to generate synthetic smoke images.A smoke detection method based on two dimensional convolution neural network is studied using these synthetic smoke images.The training of CNN requires a large number of samples,but smoke video data is lacking.In order to solve this problem,smoke is obtained by Blender simulation and real smoke extraction.Then the smoke is inserted into the for-est background images to generate the simulated synthetic smoke image dataset and real synthetic smoke image dataset.Firstly,Faster-RCNN models for the object detection are trained using the two datasets.The experimental results of testing show that Faster-RCNN models trained by the synthesis smoke image can detect smoke with high con-centration and close range.The smoke detection rate can reach more than 99%.But for smoke with low concentration and distant range,the smoke detection rate is only about 50%.Secondly,FCN models for semantic segmentation are trained using syn-thetic smoke image dataset and a real smoke dataset with a manual annotation of smoke segmentation.The experimental results show that the FCN model trained by synthesis smoke image datasets can' t segment real smoke image.The FCN model trained by real smoke dataset with manual annotation can segment some smoke target.But for the segmentation of white object such as light,the model not work.So FCN is not suitable for end-to-end smoke detection.(3)Video smoke detection method based on 3D convolutional neural network is proposed.As the Faster RCNN can only extract the spatial information from two-dimensional image,there is still false alarm and omission.The 3D convolutional neu-ral network is used to extract the inter-frame temporal information of smoke video se-quence.Firstly,a non-maximal annexation algorithm was designed to improve Faster RCNN.The suspected smoke boxes output by Faster RCNN could contain the edge of smoke and retain important movement information.They also did not overlap with each other to avoid the repeated extraction of spatial and temporal features.According to the location of the suspected smoke box on the target frame,the suspected smoke video sequences are generated.The suspected smoke video sequences are the input of the 3D convolutional neural network.Experimental results show that the 3D convo-lutional neural network significantly improves the smoke detection performance.And data augmentation methods such as changing image brightness,movement informa-tion enhancement based on optical flow method,and combination of intermediate layer features with Faster RCNN score,all these methods can improve the smoke detection performance.In addition,smoke dataset is small,deeper network has the risk of over-fitting.Deepening the network depth reduces the smoke detection rate.The width of the network affects the fusion speed of time information,and the wide network can retain more temporal information leading increase of the smoke detection rate.the recogni-tion accuracy of smoke video sequence reach 95.23%,and the false alarm rate was as low as 0.39%.A video smoke detection system is designed to evaluate and improve the smoke detection method.