Research On Video Abstract Based On Background Modelling And Attribute Learning

Surveillance cameras are widely deployed in every corner of the city with the popularity of high definition cameras and the rise of the Internet of things,as well as the slogan of ‘Peaceful City' and ‘Smart City' which is proposed by the government.The surveillance device can play a vital role in fighting against illegal and criminal and maintaining social stability.However,it comes to be a challenge for people to deal with the archive and retrieval of the massive video data.Traditional method in which we store the data directly and the retrieval the information manually has already been unable to meet the needs of large-scale video.It has become a focus of research which scholars at home and abroad are trying to solve.therefore,the research in this thesis the will be related to seeking a solution to these tough problems.We have some ideas about this issue after a number of domestic and foreign materials are studied.Then deep analyses of the research results and the existing problems is conducted.This thesis expounds the main difficulties of the current research work includes detecting the video object accurately without missing any of them,classifying the concept of foreground objects after the detection,crossing the semantic gap between the concept of object and feature descriptions during the classification,etc.On this basis,the study of video abstract based on background modelling and attribute learning is put forward in this thesis.The background modeling of the video sequence is built in use of the improved ViBe.Then the useless frames which contain no foreground objects are dropped,so that we can retain the rest frames to generate the video after the enrichment,in this way,the burden on the video file to store can be reduced;Attribute classifier is set up after the acquisition of the foreground object,using the properties study concept of the foreground object detection,by which the corresponding concept can be detected.Afterwards,the foreground object can be described by using the attribute tags.In the end,the production of abstracted video can be done in this thesis.The main content of our research is as follows:(1)The background modeling and enrichment of video sequence based on the improved ViBe is proposed.ViBe algorithm is chosen after the study of video background modeling algorithm which is fast and less memory contrast to the other mainstream method.Improved algorithm of ViBe is put forward for the disadvantage of the original ViBe in which there will be noise,flashing point and the ghost introduced in the process of initialization in practical surveillance scenario.The improvement plans are proposed respectively,they are the flash point removal method based on counting point threshold,noise eliminating method based on morphology,and the ghost area detection and suppression algorithm.the video background modeling and enrichment can be conducted after implementing the improvements of ViBe.(2)The concept detection and abstraction of video sequence based on attribute learning is proposed.Firstly,multi-kernel learning is introduced in the framework of directly attribute prediction model,then the optimization solution for the weight vector is given in this thesis.Furthermore,the proposed model is applied in video concept classification.The ability of attribute description is applied in the video summary.(3)The video abstract prototype based on background modelling and attribute learning is built with the idea of Object-Oriented Program Design after the propose of the two research points above.This system is constituted by video enrichment modules,properties prediction model training module,and video abstract module.The system runs well and meets the expected goal.