A Research On Edge Video Processing Based On DM8127

Edge computing,which offloads computing tasks to the edge of the network close to the device/data source,performs operations such as data storage and computing,and realizes flexible allocation of resources.It has the advantages of short transmission distance,low delay and high efficiency.With the development of artificial intelligence(AI)and deep learning,edge computing has been mainly applied in scenarios such as smart cities,smart transportation,and smart homes.However,edge devices have limited resources and cannot deploy complex deep learning models;and the security of video transmission has also received increasing attention.Therefore,based on the DM8127 video processing platform,this thesis conducts research on edge video processing,including:research on lightweight video target detection and encryption algorithms based on edge computing.The main research contents are as follows:1.In view of the low-latency and high-quality performance requirements of the video processing system,this thesis constructs an edge video acquisition and processing system based on DM8127.Among them,the DM8127 video processing platform is responsible for real-time processing of video data collection,compression,calculation,transmission,etc.It adopts software optimization and hardware acceleration methods to improve the operating efficiency of the system,and uses RTSP streaming media transmission protocol as the data transmission protocol for video streaming transmission,output Dual-channel video streams with resolutions of 1920*1080p and 720*480p,realize environmental monitoring functions and GUI client services.2.In view of the limited computing power of edge devices and the inability to deploy complex deep learning models,this thesis studies the lightweight algorithm of the YOLOv5 s algorithm.Referring to the structure and design criteria of ShuffleNetv2,the YOLOv5 s algorithm is lightweight and designed,v5Lite＿a,v5Lite＿b,and v5Lite＿c algorithms are proposed and tested,and the v5Lite＿c lightweight algorithm is deployed in the edge video processing system to realize the real-time target detection application of video frames.50 epochs of training were performed on the widerface data set.The results show that all three algorithms meet the lightweight requirements.The v5Lite＿c algorithm has the highest lightweight level,shortening the training time by 42% and reaching an accuracy of 0.737.After 100 epochs of training on the set,the results show that all three algorithms can accurately detect occluded target objects,and achieve mAP.5 scores of 0.99,0.994,and 0.989,respectively.3.Aiming at the high efficiency and security requirements of video transmission,this thesis proposes a video-oriented compression and encryption hybrid coding algorithm.Based on H.264 encoding and elliptic curve encryption algorithm,the video frame is compressed and encrypted at the same time.The residual macroblock of DCT transform and quantization is encrypted to eliminate pixel redundancy,and the ellipse is optimized based on the residual macroblock.The curve encryption algorithm ensures the effective mapping of plaintext pixels,and the compression and encryption hybrid coding algorithm can greatly reduce the amount of video data to be encrypted,and meet the needs of fast image transmission and security.The experimental results show that the compression and encryption coding algorithm can effectively eliminate the temporal and spatial correlation between pixels by using intra-frame prediction technology and inter-frame prediction technology,and using elliptic curve encryption algorithm to encrypt plaintext can improve the robustness of the algorithm and ensure Encryption quality to ensure secure transmission of video data.This thesis further proposes a target encryption application for the DM8127 edge video processing system,which processes video frames in real time to obtain the ROI of the region of interest,and effectively guarantees the information security of sensitive areas.