Research On Video Smoke Detection Algorithms Based On Multi-Attribute Features

In the early stage of fire,smoldering with smoke usually occurs due to the inadequate combustion.It is of great significance for early warning of fire,which is detected by smoke in early fire.As the development and the application of machine learning,the analysis and processing of smoke videos provide a new safety warning technology in the early stage of fire.Video smoke detection,which is based on computer vision,is carried on suspected smoke target via regional analysis,segmentation and feature extraction.Then,the classifier is trained to recognize the query target as smoke or not.Considering lower performances on the target segmentation and recognition of the suspected smoke,we carried out the research on the algorithms of smoke segmentation and recognition,which are described in detail as follows.(1)Aiming at solving the problems of frames with huge data and incomplete segmentation of the query target,an algorithm of smoke region segmentation is proposed based on saliency analysis in key frames.First,Cluster analysis is performed on the image color histograms,and the frames corresponding to cluster centers are taken as key frames.Considering the salient brightness feature and motion feature of the query smoke target,the nonlinear enhancement of luminance image is non-linearly enhanced and the enhanced image is used to extract optical flow feature.At last,the segmentation is realized by an criterion of main motion orientation.The experiment results demonstrate that the proposed algorithm could segment the relatively complete region to be as smoke target.(2)Aiming at the lower accuracy of smoke target recognition in complex scenes,an algorithm of smoke recognition is proposed based on the fusion of multiple features.Considering there are static and motion characteristics of smoke targets in videos,the color feature and dynamic texture feature are extracted respectively from key frames.And then,a new feature vector is constructed with the fusion of the color feature and dynamic texture feature.These vectors are used to train a classifier of ELM,and a trained ELM could be used to classify a query target.We take real videos to test the validity of the algorithm and put common environmental disturbances to demonstrate the robustness.The experiment results show that the proposed algorithm have a better recognition accuracy in a wide variety of environments.