Research Of Embedded Real-Time Video Structured System Based On Deep Learning

Video is an important source of information in people's daily life,industrial production and security,and video structure is one of the important research fields of computer vision.Video structure technology is the technology of breaking the spatial and templi dimension of video files,extracting key information for flattening,and the process of video structure,that is,the process of extracting and classifying key information across time dimensions.In the embedded environment to achieve the primary video structure function,video can be analyzed at the video acquisition end,filterout the video without analysis value of the time period and content,and filtered key information back to the back-end server,compared with the transmission of the original video,can greatly reduce the network bandwidth occupied and reduce the operation and storage pressure of the back-end server.In recent years,with the rapid development of deep learning technology,a large number of deep convolutional neural network image recognition,target detection,track tracktracking,face recognition and other algorithms have been put forward one after another,in terms of detection speed and accuracy compared with the traditional image processing algorithm has greatly improved.Similarly,with the rapid development of embedded devices and convolutional neural network acceleration devices,large deep convolutional neural networks trained on GPUs(computer graphics processor Graphics Processing Unit,GPU)can be deployed on low-power embedded devices,making it possible to deploy video structured systems on embedded platforms.This paper first analyzes the basic theory of deep learning and the composition of convolutional neural networks,and selects SSD300 as the target detection network of video structured algorithm synth sits by comparing and analyzing the algorithm principles,computations,parameter volume,accuracy and embedded device applicability of mainstream image classification networks such as Alexnet,VGG,Resnet and faster RCNN,Yolo and other mainstream target detection networks.This paper discusses the impact of different basic networks on the performance of SSD300,and finally uses MobileNet as the basic network to form MobileNet-SSD,reduces the volume of the network from 105MB to 22MB under the premise of almost no loss of precision,and constructs a special data set for the application scenario,and carries on the adaptive training to the selected network.In addition,this paper uses the method of migration learning neuron cutting,modifies the network structure of MobileNet-ssd,reduces the network volume to 63.2%under the premise of ensuring accuracy,and further reduces the number of parameters and calculations of the network.Finally,using the Intel Up Core Development Board and Neurocomputing Bar(NCS2)hardware environment,in the open VINO development framework to achieve deep learning-based embedded video structured system,the front-end camera can be transmitted back to the video screen analysis,extract the video appears in persons,vehicles and other objects,to achieve the carAdvanced property analysis,sensitive area detection,such as color and type.Experiments show that the system can realize the real-time structured analysis of 20FPS of 1080P-resolution input video source on the embedded platform,and support the transmission of test results to the back-end server.