Design And Implementation Of Distributed Space-time Trajectory Computing Platform

With the development of society,intelligent security cameras are becoming more and more popular,and the camera will use the deep learning technology to process the captured video to obtain the feature information of the target in the video.When an incident that jeopardizes public security occurs,the relevant departments usually obtain the trajectory information of the criminal target through video data,and it is time-consuming and laborious to perform the inspection by manual means.Therefore,the rapid analysis and calculation of the massive feature information,and then getting the trajectory information of the specific target in the video has important practical significance for the public security and other departments.Aiming at the above background,this thesis designs and implements a distributed computing system,which can quickly calculate the massive feature data generated by the camera and obtain various trajectory information of specific targets in the video,such as determining whether the target appears in a certain time and space range,and obtaining the trajectory of the target in a specific space-time range and the clustering of the targets in the video.The main contents of this thesis include:1.Compare the architecture of two mainstream distributed computing systems.Refer to the Apache Spark operator interface design,refer to the network interface and architecture of MPI,and combine the advantages of both to complete the design and implementation of the whole system.The advantage of the operator interface is that the developer can quickly complete the ever-changing business requirements in the trajectory scenario by combining the operator interfaces.In the system design and implementation,a simple and effective architecture model is adopted,a large number of template programming techniques are used,and an efficient memory and data management module are designed.These ensure that the system can efficiently perform various trajectory calculation tasks.2.Analyze the trajectory scene and extract three core trajectory operators,including feature clustering,feature search and companion analysis.In order to speed up the clustering,the clustering operator utilizes the spatio-temporal continuity of video data,adopts the strategy of combining and then segmenting,so that the system can quickly cluster millions of features,compared with the traditional clustering algorithm,greatly improving the speed of the algorithm.Feature search provides two methods: approximate search and exact search.The speed of approximate search is much faster than the speed of exact search,but it takes the time cost of index construction.Users can choose different methods according to the scene.Finally,the Apriori algorithm is implemented to perform the frequent item set mining.3.The function tests and performance tests of each operator in the system are carried out;the difference between the new clustering algorithm and the traditional clustering algorithm in running speed and clustering precision is compared;the performance tests of the operators shared by Spark and this system are carried out.And the advantages of this system in architecture design are explained.