Abnormal Event Video Captioning Method Of Mine Video Based On Feature Graph

At present,the research on abnormal event detection in surveillance video has become a hot issue that the society and the government pay more and more attention to.In China’s coal industry,it is an effective technical means to prevent and alleviate safety accidents that how to intelligently analyze semantic information of underground coal mine monitoring video and text description of abnormal events contained in it,so as to quickly and accurately convey safety risks.At present,due to the harsh environment,complex background,dim light and uneven distribution in underground coal mine,the video data in underground coal mine contains a lot of noise,the image quality is difficult to guarantee,and the occurrence of abnormal events is concealed,which seriously affects the accurate description of abnormal events in underground coal mine video.Therefore,this paper studies abnormal event detection and video caption methods of underground coal mine monitoring video respectively.The main research contents are as follows:(1)In order to improve the accuracy of abnormal event detection in underground coal mine surveillance video,this paper proposes a multi-classifier joint abnormal event detection method based on graph structure matching.Firstly,based on VGG16 convolutional neural network model and long short-term memory network model,feature maps based on spatio-temporal feature vectors of downhole video were constructed to represent the structural elements of feature vectors and semantic elements of video features.On this basis,the method of graph structure matching is used to reduce the interference of noise and illumination.Finally,the multi-classifier composed of random forest and support vector mechanism is used to realize the joint classification and recognition based on graph structure matching to complete the detection of abnormal events in underground coal mine video.In this paper,the underground coal mine video data set and the public data set UCSDped1 are respectively used for experimental analysis.Experimental results show that the proposed method can effectively improve the accuracy of abnormal event detection in underground coal mine surveillance video.(2)In this paper,a video caption method combining graph structure matching and relational memory kernel is proposed for abnormal events in underground coal mine surveillance video.Firstly,the VGG16 convolutional neural network model and long short-term memory network model were used to extract the temporal and spatial features of video sequences.Secondly,the graph structure is used to match and screen the spatio-temporal features to realize the structural characterization of the downhole video features and complete the video feature coding.Finally,in order to strengthen the mapping between video feature vectors and text vectors,this paper uses relational memory core RMC to decode the coding features of the above video and realize the video caption of video abnormal events.In this paper,experimental analysis is carried out on underground coal mine video data set and the public data set MSVD respectively.Experimental results show that the proposed method can accurately achieve video caption of abnormal events in underground coal mine surveillance video.