Research Of Video Similarity Analysis Method Based On Object Attribute Relationship Graph

With the development of computer technology and multimedia technology and the dramatic increase of online video,how to effectively find,retrieve and process those video data has become a problem that needs to be researched and solved urgently in the academic and industrial fields.Traditional video retrieval requires manually adding several text keywords index to the video and retrieving related videos based on the text keywords index.However,this method has the problems of individual cognitive differences,limited description capabilities,and inefficiency.In order to increase the efficiency and objectivity of video retrieval,people began to research content-based video retrieval technology.Content-based video retrieval refers to a video retrieval method in which video content is extracted and described by video content analysis automatically,and then video retrieval is performed according to similarity measurement results between video descriptions.However,the measurement of the similarity between the two videos is still a difficult problem.Therefore,this thesis conducts in-depth research on the existing methods of video similarity analysis and finds that the current video similarity analysis methods exist some problems:(1)Video is a kind of heterogeneous data with complex content.Existing video representation models mostly use vector feature matrix to describe video,and ignore the structured spatio-temporal information of video,which causes the understanding bias between underlying features and high-level semantics.(2)Video similarity analysis relies on the characteristics of video content.At present,most researches have increased the accuracy of video description by increasing the dimension of features,but it results in “dimensional disaster”.This thesis focused and studies those key issues,proposed a video similarity analysis method based on object attribute relationship graph(OARG).First,the video divided into several equally spaced frames.The Faster R-CNN object detection model is used to accurately detect the objects in each frame.The interested objects' effective trajectories in a video are dynamically tracked by a designed multi-object tracking algorithm,and then generated feature descriptors by using piecewise polynomial curve fitting and FFT.Those feature descriptors are used to construct a compact and structured object attribute relationship graph as the representation of a video.Finally,the similarity of the two video segments is measured by the graph matching method.In order to verify the feasibility and effectiveness of our method,we trained the Faster RCNN object detection model on Caffe,and coded a multi-object tracking algorithm using Python language.We successfully obtained the object effective trajectories.Finally,we do a video similarity analyzing comparison experiment of our method and the other two common methods on MATLAB.The comparison experiment results showed that the method based on object attribute relationship graph proposed in this thesis have a good effect.