Research On The Storage And Indexing Method Of Monitoring Video Information Features

With the rapid development of network and multimedia technology,video surveillance system is becoming more and more common,and it has become an important part of safe city and intelligent transportation.The video data grows exponentially and rapidly,showing massive unstructured and multi-form characteristics.At present,many research institutions and scholars have made some achievements in the field of video retrieval,but there are still some problems in the efficient representation of video information,automatic extraction of video semantic information,efficient storage of video information and efficient retrieval to solve.On the basis of the research results of video retrieval,aiming at the storage and indexing of video surveillance,this thesis presents a storage model for the monitoring video information feature and an indexing method for high-dimensional data.This thesis firstly analyzes the hierarchical structure of the video and presents a description scheme of monitoring video information feature based on the MPEG-7 standard.According to the hierarchical structure of the video,the thesis designs a storage model of the monitoring video information feature,and discusses the storage of all kinds of descriptive data in surveillance video and surveillance video in detail.Then,this thesis analyzes the dynamic operation of R-tree structure commonly used in high-dimensional index,and proposes an improved R-tree node splitting algorithm for the multi-path query problem in R-tree.By reducing the sum of the covering area of the smallest external rectangle after the node splitting,the data aggregation of the nodes is increased and the overlapping factor of the R-tree is reduced to improve the dynamic operation of the R-tree.Then,this thesis improves the writing and reading of R-tree index structure in the disk,by reducing the number of disk reading and writing in the retrieval process,which can reduce the consumption of disk I/O and improve the performance of the index.Finally,this thesis analyzes the problems existing in the process of combining Rtree with clustering method to create index.Aiming at the problem that the clustering result is sensitive to the k value of the cluster number,the initial cluster center and the isolated point,this thesis puts forward an improved clustering algorithm DKMC algorithm which is suitable for creating the R-tree.In the algorithm,the number of clusters k is determined by using the idea of hierarchical clustering algorithm and the object which is nearest to the average point that is average point of the adjacent object set is chosen as the clustering center.The proposed algorithm reduces the influence of the clustering result incompatible to the actual data distribution by setting the value of k in advance,and reduces the interference of the isolated point to the clustering result.The principal component analysis method and the DKMC algorithm are used to construct the new index structure PKMR-tree.First,the principal component analysis is used to reduce the dimension of the high-dimensional data,and then each node is regarded as a single whole to execute the DKMC algorithm starting from the root node.The divided classes are used as the child nodes of the node and each child node is divided into clusters by executing the DKMC algorithm,as a result,the tree index structure is created.The corresponding retrieval algorithm is discussed,and the experiment results show that the PKMR-tree has better retrieval performance.