Research On Video Graph Construction And Mining Based On Surveillance Video

In recent years,the field of computer vision has developed rapidly,with the rise of face detection and feature extraction methods based on deep learning.Pedestrian analysis under surveillance video has gradually become a hot research direction.Nowadays,surveillance cameras are all over the city.Mining effective information from video is of high application value for the field of urban security.In view of the pedestrian analysis problem in massive monitoring video,this paper proposes an automatic construction method of video graph,which is a knowledge base based on video.It is structured as follows,first of all,detecting and aligning pedestrians face detection in video,extracting the human face deep feature,clustering multiple pedestrians face feature,get each pedestrians many faces,the uniqueness and specify the pedestrian clustering number.A storage model based on graph database Neo4 j is designed,and Cypher language is used to add,delete,modify and check nodes and relationships.Then,according to the time and location factors,analyze and monitor the co-occurrence relationship between pedestrians in video,propose the co-occurrence relationship discovery algorithm for pedestrians,and use the front-end analytical framework to design the interactive page of video graph.Finally,useful information in video is extracted and mined based on monitoring video,including algorithms based on association rules to predict pedestrians with cooccurrence relationships.Fp-tree algorithm is used to accelerate the process of mining association rules.For large-scale data sets,Spark is used to realize distributed parallel operation algorithm.For pedestrian trajectory,it is divided into the frequent trajectory mining of single pedestrian based on suffix tree and the frequent trajectory mining of multiple pedestrians based on segmented clustering.The possible communities are detected and the algorithm of community center detection is proposed.This paper collects two real data sets and verifies the validity of the co-occurrence relation mining algorithm of video graph through experiments.The performance of the prediction algorithm based on video spectrum co-occurrence relationship is verified through the public data sets.Experiments show that the distributed Spark co-occurrence relationship prediction algorithm performs better than the traditional prediction algorithm in large-scale data sets.For algorithms such as frequent track mining,community detection and center detection,relevant experiments were carried out successively for real data sets,and the experimental results showed that the algorithm performed well.