Research On Surveillance-oriented Video Summarization

With the rapid development of information technology,the importance of intelligent video surveillance in social security is becoming increasingly outstanding.Particularly in video surveillance,huge amounts of data are recorded continuously over long periods.However,a great deal of manpower is needed to look for the key clues from videos to figure out what happened in the criminal cases.Therefore,how to summarize the video content for efficient video browsing is meaningful and need to resolved immediately.In recent years,the study of video summarization has made great progress.Based on the representation form of the video summarization,various summarization algorithms can be broadly classified into two categories: the static video abstract and the dynamic video abstract.The process of obtaining the static abstract is relatively simple,which is independent of the specific scene,while this method losses the dynamic nature of videos and generates a summary with much redundancy;the dynamic video abstract methods reduce the spatial and temporal redundancy in videos,and can achieve visually pleasing presentations,but the final results depend much on the accuracy of object detection and tracking.In video surveillance,video abstracts are usually used to help the police to search for key clues from massive videos.However,existing methods fail to provide the summary of the moving objects,which means people have to look at the details of the specific object manually.This operation is time-consuming and inefficient.In addition,attributes of the object are not utilized fully,which can improve the efficiency further.The main content of this thesis can be described as the followings.Considering the limitation that existing video abstract methods fail to summarize the video content comprehensively,a hierarchical and informative summarization framework is proposed.The video abstract obtained by this method consists of two parts: holistic-level and object-level summarization.The holistic-level summarization provides viewers with a comprehensive and compact representation of the original video,via which users can determine whether the interested obejcts exist in cuurent video.The object-level summarization summarizes the detailed information of the moving objects in the form of narrative.The two summarizations are formulated as two different energy minimization problems,which are solved by the proposed heuristic algorithms.There is a possibility that the efficiency of video browsing can be improved further.Existing methods fail to make the best of attributes of the objects,which are utilized to extract some useful information by the proposed method.To be specific,loitering indication is obtained via analysing the trajectory of the moving objects.Viewers can determine whether loitering behaviours exist directly via viewing the abstract.To improve the time performance,the proposed method is accelerated by the CUDA technology.First,the development of general-purpose computing and the CUDA framework are introduced;Second,the time-consuming modules are analysed via the experiments;third,modules are optimized via CUDA,including: background modelling,face detection,saliency detection etc;at last,the experimental results are compared and analysed.Considering the drawback of conventional video summarization system,a new system is built.During the design process,the structure of encapsulated module and interface is adopted for its advantages.Besides,the proposed method is accelerated by using the CUDA technology.The proposed hierarchical video abstract method is surveillance-oriented,which solves the problem that traditional methods fail to provide comprehensive summarization of the video content,and also improves the efficiency of video browsing further.At last,a video abstract system is built,during which the proposed method is accelerated by the CUDA.