| As an important part of a smart city,intelligent security and protection has gradually been concerned by the public,and people have put forward higher requirements for the efficiency of security in the city.With the rapidly increase of data sources in security and protection,video surveillance is usually computation-intensive and delay-sensitive.So how to optimize the delay and accuracy of video analysis is still the key to improving security efficiency.By offloading security video to mobile edge servers for calculation and analysis,the delay of video analysis has been greatly reduced.Although mobile edge computing(MEC)expands the computing power of users,it is still necessary to reasonably configure edge network resources to improve the timeliness and accuracy of video analysis due to the increasing access demand.This thesis focuses on edge offloading and resource allocation of video streams in security and protection to improve the accuracy of video edge analysis and reduce the system delay.The main researches in this thesis are as follows:(1)Aiming at the problem of edge resource allocation of the security video stream in the scenario of one MEC server and multiple users,this thesis proposes an "edge-end" collaborative video stream offloading and resource allocation architecture based on MEC.In order to improve the accuracy of video edge analysis and reduce the system delay,this thesis describes the process as a cost minimization problem composed of delay and video analysis accuracy.To solve this problem,this thesis designs an optimization algorithm of joint video frame resolution scaling,task offloading,and resource allocation based on deep Q network(JVFRS-TO-RA-DQN).The algorithm innovatively contains two DQN networks for training.One DQN network is used to train and select the action of video stream offloading and resource allocation,and the other is used to train and select the action of video frame resolution scaling,then the two actions are combined to speed up the convergence process.Simulation results show that,compared with other baselines,the JVFRS-TO-RA-DQN algorithm can better optimize the offloading strategy,reduce the video stream edge analysis delay,and improve the accuracy of video analysis.(2)Due to the limited communication resource and computing resources that the MEC server can provide,when a large number of computationally intensive tasks occur in the network(such as complex image processing,video stream segmentation,multiperson pose recognition,etc.),only one MEC server is not enough to efficiently complete the calculation,and the assistance of cloud server is needed.Aiming at the video edge offloading and resource allocation problem in the cloud-edge collaboration scenario with multi-MEC servers and multi-terminal devices,this thesis proposes a three-layer optimization model of "cloud-edge-end" based on MEC considering the competition between MEC servers and between cloud servers and MECs.On the premise of ensuring the accuracy of video analysis,the optimization model is established to minimize the total system delay.The DQN algorithm is trained with the random strategy,when the spatial dimension of the action is too large,the sample size will increase and its convergence performance will decline.To solve this problem,a cloud-edge cooperation algorithm based on deep deterministic policy gradient(CE-DDPG)is proposed in this thesis.The algorithm uses DDPG to realize the high dimensional continuity of state space and action space in MEC scenes.Based on this,the empirical replay mechanism is improved to make efficient use of the empirical samples in the algorithm training stage.Simulation results show that,compared with other baselines,the CE-DDPG algorithm has faster convergence and better stability,and better optimization effect for reducing video stream edge analysis delay.Meanwhile,it is verified that under the same system parameters,the video analysis model based on cloud edge collaboration has better performance than other baseline methods. |
PDF Full Text Request