Research And Application Of Edge Cloud Resource Scheduling Technology For Video Analysis Task

With the widespread application of object detection,image classification,anomaly detection and other technologies in the field of video surveillance,the demand of real-time video analysis is increasing.Due to the explosive growth of video data,the video surveillance system based on cloud data center is confronted with high cost of network bandwidth and low timeliness of data.As a promising computing paradigm,edge computing can provide low latency video analysis service close to the video source at the edge of network.However,since the computing resource of the edge cloud is usually limited,it's a challenging problem to meet low latency and high accuracy requirements of all users,especially in the case of a surge in the number of requests,which poses a problem to balance the accuracy of video analysis tasks with the overall throughput of the system.This thesis focuses on the edge cloud resource scheduling strategy for video analysis tasks.Firstly,in order to solve the trade-off between task accuracy and system resource utilization,the video quality,delay requirement and computing resource constraints of the video analysis task are analyzed and a fine-grained mapping model of computing resource and task accuracy is constructed.On this basis,this thesis designs a video quality and computing resource offline allocation algorithm based on the divide-and-conquer idea.Considering that the empirical values of optimal configuration of various tasks in edge scenes are unknown,this thesis proposes a gradient-aware method for online configuration of video quality and computing resource,which learns the optimal allocation strategy with the aim of maximizing the number of accurate response tasks in the long run.Secondly,in the deployment phase of the task instance,considering the interference of edge environment factors on the algorithm performance,in order to ensure the processing accuracy of the video analysis task further,this thesis proposes a video analysis algorithm selection strategy based on the comparison of the similarity between video frames.In order to reduce the network overhead caused by the upload of video clips and improve the computational efficiency,this strategy adopts a sampling mechanism based on the combination of inter-frame difference and background difference to preprocess the video clips.Then the strategy defines the algorithm selection problem as a video source similarity comparison problem.By calculating the core distance of key frames,it selects the best video analysis algorithm that consumes low resource and satisfies the accuracy constraints adaptively for the video analysis task at different moments and regions.Finally,this thesis designs an edge cloud platform for video analysis tasks,which contains an online resource scheduler and an adaptive algorithm selector,and provides users with basic capabilities such as video access management and streaming media management.Experiments show that the methods proposed in this thesis can meet the accuracy requirements of users in a variety of edge scenarios,and improve the throughput of accurate response tasks significantly.