Research On Fire Video Region Detection Based On Depth Temporal And Spatial Features

With the rapid development of society and economy,the factors that cause fires are increasing,and various fire accidents continue to cause serious threats to the safety of people’s lives and property.Therefore,rapid fire warning has always been a key issue of domestic and foreign concern.Fire video surveillance systems have gradually become popular.Because flames are an important visual feature of fires,the flame region detection method based on video images has become an important auxiliary method for fire detection and rescue,especially the dazzling performance of deep learning in various fields and electronics.With the rapid improvement of hardware computing power,some researchers are also committed to studying fire video flame images based on deep learning.However,there are few researches on flame region detection based on deep learning.Considering that the convolutional neural network used for flame region detection has a complex structure and is easily disturbed by complex backgrounds,this paper adopts a two-stage detection step to realize the recognition of video flame region.The first stage of detection is to use an improved ViBe algorithm to extract video clips containing suspected flame regions.In the second stage,a dual-stream structured flame video region detection network based on deep spacetime features is proposed to detect flame regions in suspected flame video clips.The main research contents of this paper are as follows:(1)This paper improves the ViBe algorithm and uses it to filter video clips that do not contain the foreground.Aiming at the defect that the current multiple foreground detection algorithms are too corrosive to the flame region,this paper designs a new dynamic background detection rule based on the characteristics of flashing pixels and flame color,and improves the background template update mechanism of the ViBe algorithm on this basis.In order to retain a more complete region of suspected flames,and to suppress the interference of dynamic background.Through experimental comparison,it is proved that the improved method proposed in this paper is more suitable for the field of flame detection.(2)In this paper,the method of pre-dividing pixel blocks is used to construct a lightweight flame area detection network CNNFlame.In view of the high false detection rate and complex structure of the current deep convolutional network model in the flame detection task,this paper adopts the idea of constructing a lightweight flame image area detection network and then adding a time stream to extract the dynamic features of the video.Based on the similar characteristics of the flame local features and the overall features,this paper divides the input video frame into grids of equal size as candidate areas,and then plans the size of the output feature map according to the grid size,thereby constructing a lightweight flame area detection network Frame CNNFlame,check out all grid areas at once.Experiments show that the detection accuracy of CNNFlame is slightly lower than that of FCN,but it greatly reduces the complexity of the network,which is conducive to real-time detection and subsequent improvements.(3)This paper builds a two-stream flame region detection network TSCNNFlame on the basis of CNNFlame.Since the pre-divided candidate regions of CNNFlame cannot adapt to flame blocks of different sizes and shooting distances,this paper introduces SK convolution to enable the network to adaptively adjust the size of the receptive field to improve network performance.In order to suppress the increase in parameters caused by SK convolution and further lighten the network,this paper compares and analyzes the feasibility and parameter compression rate of the three kinds of lightweight modules that have been prominent in recent years and the SK convolution fusion,and finally chooses ShuffleNetV2 convolution The SK-shuffle module after the fusion of the module and the SK convolution reconstructs the TSCNNFlame network.Through the multi-directional comparison with the current method in the experiment,the high precision,anti-interference and real-time performance of TSCNNFlame are verified.