Design And Implementation Of Image Recognition Technology Based On Evidia Embedded System

With the development of science and technology,traffic gradually tends to be intelligently managed,and the detection of accidents on highways has become a hot issue.Image recognition technology can identify and mark targets in the image that match the type of object being trained It is less affected by image background interference,and can be transplanted to embedded development platforms.It has the characteristics of fast,accurate and adaptable,and has important significance in the application of urban and rural roads.In this project,an image recognition algorithm based on Nvidia's embedded system is designed to complete the rapid and accurate identification of specific targets on the highway.First,the image recognition algorithm based on target detection is studied.The current image recognition algorithm is based on the convolutional neural network.The input layer performs feature extraction at the convolutional layer,simplifies feature data at the pooling layer,increases the nonlinear correlation between features at the activation layer,and sums multiple features at the fully connected layer.After analyzing and comparing the current mainstream image recognition algorithms,the YOLO v3 algorithm is selected.The YOLO v3 algorithm is based on the darknet-53 network structure.It can output recognition results at three scales and continuously optimize the model through the loss function.The detection recognition rate reaches 80% and the rate is 0.326 seconds.Then designed the software of image recognition system.A large number of image data sets become the training set after screening and target annotation,and the model training is performed to obtain the weight file.After the weight file is imported into the algorithm structure,the image to be detected is identified,and the obtained identification result is marked on the image.In order to further reduce the demand for high-performance hardware,the YOLO v3 algorithm is tailored and optimized to obtain the YOLO v3-tiny algorithm,which is suitable for embedded development platforms.Finally,the interaction and test analysis of the image recognition system is carried out.Through the network,the image data is exchanged among the monitoring terminal,the embedded development platform and the computer.The computer develops the upper computer software which displays the image data recognized by the embedded development platform in real time,records the received time node,stores the image data,and can view the historical information of the previously received image.