Research On Smoke Detection Algorithm Based On Surveillance Video In Field Scene

There are many forest fires every year around the world.Forest fires not only bring huge economic losses,but also seriously threaten our ecosystem.Once the forest is on fire,it is difficult to extinguish it because of terrain and climate.Therefore,it is important to detect forest fires early.In recent years,video surveillance has become the most extensive surveillance tool in China,and it has been widely used in law enforcement,security,and ecological environments.Fire smoke detection based on surveillance video has low cost,high efficiency,wide coverage,and early detection of fire.Due to the intelligence,real-time and accuracy of image-based and video-based methods,it has received more and more attention from researchers.This paper focuses on the field smoke detection,using traditional and deep fusion methods and semantic segmentation for smoke detection.The main work is as follows:(1)In this paper,the video frame is first filtered;then the color space and motion detection are used to preprocess the smoke area to extract the color and motion features of the smoke in the video;finally,the fusion features are classified using SVM and deep network respectively.It compares the effects of two motion detection algorithms and two color spaces in smoke detection,and obtains a relatively good detection effect through different combinations.However,due to the shallow network,the learning ability in processing high-definition pictures is not enough;there is too much preprocessing.Compared with the end-to-end network,there are too many human interference factors,many steps,and not simple enough,which will also affect the accuracy to a certain extent.(2)Aiming at the interference of human factors in the traditional and deep integration algorithms,this paper proposes an end-to-end semantic segmentation network with small model parameters but strong network learning ability and does not require any pre-processing and back-end processing.In the experiment in this paper,we compare three classic and advanced networks in semantic segmentation.The network proposed in this paper is more advanced and better in smoke detection,but smoke detection based on image semantic segmentation will ignore the motion information of smoke,which will affect the accuracy and false detection rate of detection,so this method is still flawed(3)Considering the shortcomings of the above methods,this paper uses a three-dimensional convolutional neural network to detect smoke in video.We propose a more effective 3D neural network architecture for video semantic segmentation in wildfire smoke scenes.The model uses the encoder-decoder structure based on ResNet.In the encoder stage,we use 3D residual blocks to extract the spatiotemporal features of smoke.The decoder upsamples the output of the encoder three consecutive times.Then pass them to the three smoke graph prediction modules respectively,and supervise the output smoke prediction graph through the binary image label,and finally obtain the final prediction through feature map fusion.The model can achieve end-to-end and pixel-to-pixel training without pre-training from scratch.It effectively overcomes that the two-dimensional convolutional neural network cannot detect motion information and does not require pre-processing,and there is little manual intervention.Experiments show that this method is more effective than the above method,but it still has deficiencies in detecting small targets with slow movement and long distance.