Video Monitoring Ststem Based On Edge Computing And Microservice

In 5g network,the emergence of mobile edge computing makes the computing tasks of the terminal can be unloaded to the high-performance server on the edge side for processing,which greatly improves the performance of various applications and shortens the service transmission delay.Among them,video monitoring is the most widely used scene in edge computing.The existing intelligent monitoring system has serious problems in video signal processing.The computing and transmission bandwidth load of video monitoring in traditional cloud mode is heavy.Edge computing can realize the localization of video data in the process of transmission stream,which brings the advantages of low delay and efficient resource utilization for video monitoring system.However,there are still two problems:on the one hand,most of the related researches at home and abroad focus on the design of computer vision algorithm model or overall framework,and still stay in the simulation or demo stage,without enough convincing evidence In order to understand the necessity and value of edge computing,the ideal way is to actually build a set of IOT system and deploy business algorithm to compare the effect.On the other hand,due to the complex functions of the video monitoring system,it is necessary to add a variety of algorithm modules with rich services in the future,but it is difficult to expand the traditional architecture.With the development of virtualization container technology,we can deploy a variety of algorithm modules separately,which greatly reduces the difficulty of system expansion.The huge micro service system is composed of multiple containers.If a single module fails,the whole project will run abnormally.However,it is very difficult to locate the fault module,because the traditional one-step debugging method is not suitable for the microservice architecture.In order to improve the stability of the whole microservice system,it is necessary to design an efficient fault detection and root cause location algorithm.Based on the above two problems,this paper builds a video monitoring system based on edge computing,and designs a micro service attribution algorithm to quickly find system faults and locate the source of faults.The main contributions of this paper include the following three aspects1.Design and implementation of edge network platform:This paper uses the relevant equipment of amarisoft to build the traditional cellular network platform,uses open vswitch to complete the deployment of MEC,realizes the basic functions of edge server,and tests the function and performance of edge computing platform through 1080p video transmission task.In general,the edge computing platform built in this paper effectively realizes the characteristics of low latency and high speed of edge computing,which provides the basis for the subsequent business building and algorithm simulation.2.Construction of video monitoring system:Based on the physical platform of edge computing,this paper designs several algorithm models according to the characteristics of edge service scene,builds the actual business of video monitoring,and fully proves the application value of edge computing.Firstly,GMG algorithm is used to detect effective video clips at the camera terminal,and H264 is used to compress the video.After testing,the accuracy rate of the combined model is 36.5%,the recall rate is 95.7%,and the total volume of filtered compressed video clips is only 92.4%of the total video clips,which greatly reduces the bandwidth pressure.Secondly,transfer learning method is used in the edge side to fix the parameters of several layers of RESNET neural network under multi classification targets and train with finetune.The network parameters before softmax function are stored in the database as vector index.Face recognition can quickly retrieve in the database in real time.The average delay of tasks can be reduced by 80.6%when the cloud task is brought to the edge side.Finally,in order to improve the accuracy of the model under the condition of ensuring the edge data privacy,this paper introduces the federal average training method and network distillation technology,which improves the accuracy of the edge side face recognition model by 15.3%.3.Microservice system monitoring and research:in order to better expand the functions of the platform in the future,each model is deployed and run in a process in the form of service,giving the system flexible code organization and release rhythm.Microservice architecture not only brings convenience to development,but also brings challenges to system reliability.The monitoring of system status and the fast tracking of system failure are important issues for enterprise platform.The first part is anomaly detection.In this paper,we preprocess the data by box line graph and linear interpolation.We use DTW algorithm to remove the phase deviation and calculate the time series similarity.Then we extract the time series features by K-means clustering and Tsfresh tools,and detect the anomaly based on logistic regression and random forest algorithm.The second part is the root cause localization.After the failure probability of the time sequence point is obtained,a directed and weighted fault call graph is constructed by using the custom statistical strategy,and a new random-walk algorithm is designed to realize the root cause localization by adding self-call and inverse-call edges to PageRank.Finally,based on Zipkin and Prometheus,this paper monitors the call link level and machine index level of the microservice system,stores the results in Elasticsearch and other databases,visualizes them through Grafana,and injects faults by using Istio and other tools.On the fault data set,the average time of the proposed algorithm is 0.523 s,and the accuracy of root cause ranking in top 1 and top 2 is 84%and 94%respectively.Compared with Google's classic PageRank algorithm,the accuracy is improved by 23.5%and 16.1%.