Dual Channel Video Flame Detection Method Based On Deep Learning

Among many disasters,fire occurs most frequently,which brings economic losses and casualties to human society.Once a fire occurs,the fire will spread rapidly,so it is necessary to carry out effective detection and alarm at the beginning of the fire.The traditional smoke and temperature detection technology has become mature,but with the rapid development of economy and the continuous advancement of urbanization,this kind of traditional contact fire detector can no longer meet the detection needs of large space places.Video fire detection technology has the advantages of non-contact,fast response,comprehensive fire information and high cost performance.It can not only meet the needs of large space places,but also realize faster and more accurate fire detection.In recent years,the rise of deep learning has brought new development to video fire detection.The video fire detection based on deep learning can automatically extract the color,shape,texture and other features of flame,so as to achieve more accurate detection than the traditional video fire detection technology with manual feature extraction.Flame has high temperature characteristics,so the introduction of infrared technology also makes video fire detection technology have better anti-interference ability.At present,the video flame detection based on deep learning lacks the fusion of infrared information and visible light information,and the research on the flame scale characteristics in the early stage of fire also needs to be developed.Based on this,this paper proposes a dual channel video flame detection method based on deep learning,which comprehensively utilizes the flame scale information,visible light image information and infrared image information to improve the flame detection performance.The specific research contents are as follows:1.The adaptability of the out-of-state object detection algorithm for flame detection is compared and tested to find the best flame detection algorithm.The standard data set of flame detection is established by means of public data set collection and self built collection platform for shooting.The five out-of-state object object detection algorithms:Yolov3,Yolov4,Efficientdet,SSD and Faster R-CNN are selected for training and testing in the established data set.The test results show that Yolov3 object detection algorithm has achieved the best experimental performance in the flame test set,and the P-R curve can fully surround the other algorithms.The AP value,accuracy,recall and F1 score can reach 83.3%,83.8%,84.0%and 83.9%respectively.2.The multi-scale convolution module and two-way feature fusion structure are constructed,and a multi-scale video flame detection algorithm is proposed to enhance the flame detection response ability in the early stage of fire.The standard data set of early flame detection is established by building an early flame image acquisition platform.Based on the performance test results of object detection algorithm,a multi-scale video flame detection algorithm for small-scale flame is designed.In the backbone network stage,a multi-scale convolution module is designed.In the module,the structure of block convolution and dilated convolution is adopted to extract the features of the picture at different scales,so as to improve the sensitivity to small-scale flames.In the feature extraction stage,a two-way feature fusion method is designed to extract the effective feature layer from bottom to top and from top to bottom,including feature extraction and fusion at different scales.The algorithm is trained and tested in the established early flame detection standard data set,and achieves better experimental results than yolov3 object detection algorithm.The accuracy,recall,accuracy,false positive rate and AP value are 98.7%,93.7%,96.3%,1.2%and 93.2%respectively.At the same time,the number of parameters,storage space and floating-point calculation of multiscale video flame detection algorithm are less than yolov3 object detection algorithm,and the prediction time of a single picture is only 0.08s,which has more advantages in practical engineering application.3.A dual channel video flame detection algorithm based on deep learning with UAV as the image acquisition front end is proposed.In order to avoid the influence of light source,heat source and moving object on flame detection in the actual scene,a flame detection algorithm with two image input channels of visible light and infrared is designed to form a richer feature layer.The effective feature layer of visible light channel and infrared channel is fused in the feature extraction stage for flame classification and location.Take UAV as the image acquisition front end,build a dual channel flame video image acquisition platform,and establish a dual channel flame detection standard data set based on UAV.Based on this data set,the dual channel video fire detection algorithm is trained and tested.Compared with the single channel fire detection algorithm,the dual channel video fire detection algorithm has achieved better experimental results.The accuracy,recall,accuracy,false positive rate and AP value are 98.9%,98.8%,98.7%,1.4%and 98.2%,respectively,which have benefited.Among them,the improvement of recall rate is particularly obvious,In the test process,it shows the improvement of flame recognition ability under the interference of light source and shelter.4.Establish a comprehensive flame detection data set.In the process of developing the algorithm,by building a flame video image acquisition platform,continuously expand the flame data set,establish a comprehensive flame detection data set with rich samples and clear research pertinence,which includes a variety of flame scales,combustibles,combustion background,distractors and flame images of image acquisition front-end equipment(video surveillance camera and UAV Airborne Camera),At the same time,it is also the only flame detection data set including dual channel flame images corresponding to the time and space of visible light channel and infrared channel.After sorting,it will be published on the public website for researchers to obtain freely,so as to provide data support for future video flame detection research work.Some data set images are shown in the complementary.