Research On In-Vehicle Video Process Strategy Based On Cloud-Edge Collaboration

With the continuous improvement of transportation infrastructure,the number of domestic motor vehicles and drivers continues to climb.However,due to the surge in the number of motor vehicles and the lack of driver awareness of safe driving,the number of traffic accidents remains high.In-vehicle video surveillance system can monitor driver behavior through in-vehicle cameras,use behavior recognition algorithms to realize the intelligence of the monitoring system,and can provide real-time warning of drivers’ dangerous driving behavior,and use video data to restore the accident scene after the accident occurs.However,the invehicle scenario is different from other fixed scenarios with random locations and variable environments.The traditional video surveillance system based on the centralized deployment model of cloud center has poor real-time performance and high bandwidth requirements,while edge computing technology can sink some latency-sensitive tasks to the edge side,so it is of great practical significance to design and deploy a collaborative cloud-edge video surveillance system for in-vehicle scenarios.In this thesis,we design and implement an in-vehicle video surveillance system based on cloud-edge collaboration by studying the full-link video stream processing and transmission strategies in in-vehicle scenarios.The system adopts Cloud-Edge collaboration architecture and containerized cluster deployment method,which can realize intelligent analysis,reversible scaling and client-side application of in-vehicle video.The main contents of this thesis are as follows.1.The cloud-edge collaborative in-vehicle video surveillance system is designed.The functional structure of the system,the cloud-edge system architecture of edge node devices and cloud servers,and the user-oriented client architecture are designed for the large volume of video data and the high time-sensitivity of intelligent algorithms.A Kubernetes cluster is built and a Kafka message queue is used to realize the communication between cloud edge nodes,while a Flink module is deployed in the cloud center node for real-time video streaming computation.2.To address the latency sensitivity of the behavior recognition algorithm,this thesis decomposes the behavior recognition algorithm into two modules,image behavior recognition and video behavior recognition,which are deployed at the edge nodes and the cloud center,respectively,to determine whether the behavior in the video belongs to dangerous driving behavior and the specific dangerous behavior category,respectively.The behavior recognition accuracy of the two modules reaches 99.09%and 83.33%,respectively.3.To address the problems of transmission bandwidth and large storage space of high-resolution video,video compression and recovery algorithms are deployed in the cloud edge nodes,respectively,and the use of reversible neural networks in video scaling is studied,and the network coupling layer is redesigned to achieve reversible scaling of video.The compressed video occupies 57.73%of the original space,the average PSNR of the recovered video and the original video is 29.68 dB,and the average SSIM is 0.9854.4.The complete function test and performance test of the in-vehicle video surveillance system implemented in this thesis were conducted,and the communication and function calls between each node passed the test.The test results show that this system can load a certain number of vehicle video surveillance terminals for processing and analysis.