Design And Optimization Of Mobile Edge Video Surveillance Systems Based On Reinforcement Learning

Video surveillance system is an important application of the Internet of Things(Io T).The huge amount of Io T data it generates and the users’ Quality of Service(Qo S)requirements for low latency and intelligent system bring great challenges to the whole system.In recent years,‘Edge+ AI’ has become an option to intelligently address massive data processing and is widely used.It can deploy the computation ability of the cloud to the edge of the video surveillance system,which can well support computation-intensive and latency-critical tasks.On the other hand,many Artificial Intelligence(AI)algorithms are beginning to be applied for image recognition,image compression,computation offloading,resource allocation,and other areas,using deep neural networks to make intelligent decisions for tasks.In this thesis,taking face recognition application as an example,we first study how to build video surveillance systems based on Mobile Edge Computing(MEC),then design and optimize the task computation offloading and communication resource allocation based on the Reinforcement Learning(RL)algorithms.Then,we study how to build an Open Radio Access Network(ORAN)video surveillance platform that supports video capture and data transmission.First,we study the design and optimization of joint decision algorithms for recognition task computation offloading and image compression rate selection for a single-camera video surveillance scenario.In order to achieve high recognition accuracy and low recognition latency of the system,we design the image recognition algorithm and the computation offloading strategy.On the one hand,we adopt Local Binary Pattern(LBP)recognition algorithm with low complexity at the front-end camera sensor and the Convolutional Neural Network(CNN)recognition algorithm with high complexity at the back-end MEC server,respectively,to complete the recognition tasks.On the other hand,we adopt the action-value method and the -greedy algorithm as the system’s decision making module for jointly optimizing the offloading strategy and image compression rate.Simulation results show that the action-value method can improve the efficiency of the recognition task computation offloading,and obtain higher recognition accuracy and lower process latency than other strategies;the -greedy algorithm can adaptively adjust the decision strategy and obtain better performance when the communication environment conditions vary.Then,we study the design and optimization of joint decision algorithms for recognition task computation offloading,channel resource allocation,and image compression rate selection for a multi-camera video surveillance scenario.We utilize reinforcement learning algorithm,i.e.,Deep Q-Network(DQN)algorithm,as a decision algorithm to address the impact of the curse of dimensionality on traditional iterative methods.In addition,to reduce the training parameters thus saving computation and storage resources,we propose a two-layer hierarchical learning framework,i.e.,DQN and Layers based on Back Propagation Neural Network(DQN+NN)algorithm.Simulation results show that both proposed RL methods optimize the decision performance,and the DQN+NN algorithm outperforms the other strategies in terms of both convergence speed and total reward value.Finally,to support the data transmission for video,image and other tasks,we build an ORAN platform for video surveillance systems.We use Open Air Interface(OAI)to build the core network(Evolved Packet Core(EPC))and 4G base stations(evolved Node B(e NB)),and use USRP B210 to transmit and receive signals.We compile and test each part of the ORAN platform to verify the connectivity of the platform.Then we perform data transmission rate test.In addition,we build the ORAN platform by connecting Raspberry Pi 4B and USB camera to capture the video.We program the Raspberry Pi with Open CV to send back the surveillance video in real time and display it on the remote server terminal.