Research On Methods For Video Image Based Forest Fire Smoke Detection

Forest fire prevention is an important topic of protecting lives and property safety,and the early fire smoke onset should be an important stage for fire alarm,smoke detection can help to find fire events earlier or prevent it from eruption.Therefore,this topic mainly focuses on the detection of forest fire s moke in the early stage of forest fire.Unlike indoor fire smoke detection,there are many influence factors in forest,such as long distance,changing light conditions and the great influence of the weather,the operation of actual detection needs to overcome the complex environment.This paper focuses on the key problems involved in forest fire smoke detection.At present,the forest fire detection system can adopt the front-end identification method or the back-end(central)identification method.The recognition module of the fron t-end is placed in the front-end.Its advantage is that it can guarantee the video quality and is not affected by network bandwidth.However,the front-end equipment module does not have high-performance operating conditions,so that it cannot directly run deep learning algorithms.The back-end recognition mode with analysis and recognition module is located in the central video analysis server,which can take advantage of the strong processing capacity of the central server to perform high-performance of video analysis and identification.Considering the software and hardware equipment conditions in engineering application and different requirements for system maintenance and transformation,this paper applies traditional method and the deep learning method for designing and implementation of the algorithms respectively.The main research content of this paper is described as follows:(1)A motion region extraction method based on motion and lightness saliency is proposed.The emission of moving fire smoke can create disconnected regions,moreover,due to background,refraction and other factors,there are differences in saliency among local regions,so we should calculate the local salient values of each connected motion region.In addition,for dynamic and static saliency feature extraction of each connected motion region,on the one side,since fire smoke usually move slowly,the motion saliency of adjacent pixels in connected region is relatively weak,which is called the weak dynamic salient region.On the other side,its brightness features are more salient than color,texture,contour,etc.For these reasons,this paper proposes a motion region extraction method based on the motion and brightness saliency in the forest fire smoke detection task.First,a global motion saliency method is proposed to extract connected motion regions,then,in each connected motion region,combining spatial information and using Gaussian model,the regional features of weak dynamic saliency are retained.On the other hand,a computing model of brightness saliency value is built in the Lab color space.Finally,the group sparsity coefficients are calculated based on the motion saliency and brightness saliency values in the global saliency region,and the suspected fire smoke salient regions are obtained based on the group sparsity ROSL algorithm.This method further reduces the disturbance of non-significant motion regions in the forest environment and obtains more accurate salient regions,thus reducing the false alarm rate and improving the efficiency of forest fire smoke detection.(2)A salient motion region extraction method based on full convolutional neural network is proposed.Currently,there are two main strategies to extract the salient motion regions of video images based on deep learning.One is to use the deep learning framework of convolutional neural network combined with multiple contexts to learn the depth characteristics.This method over-extracts the deep feature information of the image super pixel and ignores a part of the semantic information within the block,resulting in insufficient pixel position information.The other strategy is to train the model on the multi-frame dense optical flow,which is expensive to calculate and cannot meet the real-time requirements of target positioning.Moreover,the adjusted model is only targeted at specific objects.Considering that full convolutional neural network can directly generate image pixel-level mapping and is widely used in image segmentation,this paper combines the semantic segmentation function of full convolutional neural network with the salient region extraction target.Firstly,a salient region extraction model based on full convolutional neural network is proposed to construct the loss function of network training,and the pixel-level saliency mapping of static images is directly obtained.On the other side,this paper combines static and dynamic saliency detection method,extends the end-to-end static salient map to the prediction of dynamic saliency,and uses the dynamic saliency model based on the full convolutional neural network with video sequences,the static salient map as the input,can help to modify salient motion region of the video and the salient moving objects can be obtained quickly and accurately.The final experimental results show that the method in this paper can quickly and effectively obtain the salient motion region in the complex environment of the forest fire smoke detection.(3)A sparse feature extraction-based forest fire smoke recognition method is proposed.Conventional objects recognition algorithms directly extract the features of the dynamic region,while ignoring the characteristics such as fire smoke translucency,variable thickness and divergence,and fire smoke components often fuse with background information.Therefore,this paper constructed an elastic-net and distinctiveness constraint based fire smoke region separation model,the components of local area are treated as the composition as a linear combination of the background and foreground,and the classification and recognition are conducted by treating the sparse coefficient vectors as the fire smoke features,improves the discrimination of smoke and non-smoke coefficients characteristics.In terms of the design of classifier,it is considered that Ada Boost can enhance the classifier by improving the classification performance of base classifiers.However,the traditional Ada Boost is not robust enough to improve the performance of base classifiers,and cannot enhance the base classifier with "strong" performance.In order to solve the adaptive parameter selection and weight updating problem of Ada Boost,a robust Ada Boost(RAB)method is proposed in this paper.Experimental result demonstrates that the separation of smoke and background components and extracting the sparse features can provide reliable information of pure smoke for subsequent fire smoke feature extraction and recognition,and the improved RAB-RBFSVM recognition algorithm is proposed to obtain high ro bustness and accuracy of classification.(4)A P3D-DenseNet based forest fire smoke detection method is proposed.For forest fire smoke recognition based on deep learning,the 3D convolutional neural network(C3D)proposed in the early stage can capture motion behavior and static morphological characteristics from both temporal and spatial dimension at the same time.However,C3 D network has problems such as complex computation and large memory consumption.In this paper,referring to pseudo-3D(P3D)network,the convolution operation of learning the 3D information of images is replaced by 2D space convolution and one-dimensional time convolution to simulate 3D convolution,which can not only be well used for video feature expression,but also greatly reduce the computation.However,P3 D network integrates this design into a deep residual learning framework,and with the deepening of the network,such module cascade has problems of model degradation and gradient disappearance.On the other hand,P3 D network connection adopts Res Net structure,which cannot make full use of network characteristics of each layer.Therefore,in order to further optimize P3 D network,this paper introduced DenseNet to reconstruct P3 D module,realize the reuse of each layer of network features,and different fusion P3 D module formed P3 D Dense Block,keep the diversity of structure Block,while it can solve the problem of computational complexity,further alleviate the problems of gradient disappearance and model degradation in network training,and the efficiency and accuracy of forest fire smoke recognition based on continuous frame video image are improved.