Research On Fire Detection Algorithm Based On Computer Vision

Fire is a serious natural disaster.Early detection and prevention of fire are of great significance to protect people’s lives,property,social economy and natural resources.This is a research topic with both theoretical and practical research value.Recently,with the popularization of surveillance cameras and the continuous development of image processing technology,fire detection algorithms based on computer vision have been widely researched and applied.However,the current fire detection algorithms based on computer vision generally have the problem of poor adaptability to different environments.This paper conducts in-depth research on the problems of the existing fire detection algorithms based on computer vision:Firstly,in order to take full advantage of the complementary among different flame features and the adaptability of flame features to different classifiers,the various features of the flame and the relationship between the features and the classifiers are tested in this thesis.A variety of static and dynamic features of the flame are tested and three types of features are selected with strong complementary among the test results,which are color,texture and shape change features.Then the features are recognized by using different classifiers,including Support Vector Machine,K-Nearest Neighbor,Decision Tree and Random Forest.Through testing,it is found that different classifiers achieve different recognition results for the same flame features,so it is concluded that the features have different adaptability to different classifiers.Secondly,according to the above conclusions,i.e.different features are complementary and there are different adaptability between features and classifiers,this paper proposes a fire detection algorithm based on improvedDS evidence theory multi-feature multi-classifier fusion is proposed.The proposed algorithm first obtain the candidate flame area by using the Background Subtractor K-Nearest,and extracts three features of color,texture and shape change from the candidate area.Then,the three features are classified with Support Vector Machine,K-Nearest Neighbor,Decision Tree and Random Forest classifiers respectively,and the four classification results corresponding to each feature are fused using improved DS evidence theory.Finally the fusion results of the three features are fused again to obtain the final result.The algorithm is a high-level decision-level fusion method,which makes full use of the complementary among the recognition results of different classifiers.It can meet the effectiveness and stability of fire detection under different scene changes,and can ensure the accuracy of detection.Finally,the flame detection algorithm based on deep learning has some problems due to the complexity of the model,such as the high requirement of the hardware and the large amount of calculation.This paper proposes a fire detection algorithm based on the combination of depthwise separable convolution and YOLOv3.The algorithm first uses the depthwise separable convolutional neural network to classify fire images.The depthwise separable convolutional neural network can save a lot of detection time under the premise of ensuring classification accuracy.Then the identified fire images is detected by using YOLOv3’s target regression function to locate the fire location,avoiding the problem that YOLOv3’s simultaneous classification and position regression will cause the detection accuracy cannot be guaranteed.Compared with traditional video-based fire detection methods,the algorithm greatly improves the accuracy of fire detection.Compared with some standard deep learning network models,there are no special hardware requirements.The proposed method greatly reduces the amount of calculations and parameters and the detection rate and detection accuracy have been significantly improved.