This example shows how to deploy a Simulink® model on the NVIDIA® Jetson TX2 board for classifying webcam images. This example classifies images from a webcam in real-time by using the pretrained deep convolutional neural network,
ResNet-50. The Simulink model in the example uses the camera and display blocks from the MATLAB® Coder™ Support Package for NVIDIA Jetson and NVIDIA DRIVE Platforms to capture the live video stream from a webcam and display the prediction results on a monitor connected to the Jetson platform.
Target Board Requirements
NVIDIA Jetson embedded platform.
Ethernet crossover cable to connect the target board and host PC. (if you cannot connect the target board to a local network)
USB webcam connected to the USB host port of the target.
A monitor connected to the display port of the target.
V4L2 and SDL (v1.2) libraries on the target.
GStreamer libraries on the target.
NVIDIA CUDA® toolkit and driver.
NVIDIA cuDNN library on the target.
Environment variables on the target for the compilers and libraries. For more information, see Install and Setup Prerequisites for NVIDIA Boards.
Development Host Requirements
GPU Coder™ for CUDA code generation. For a tutorial, see Get Started with GPU Coder (GPU Coder).
Deep Learning Toolbox™
Computer Vision Toolbox™
Image Processing Toolbox™
GPU Coder Interface for Deep Learning Libraries support package.
Deep Learning Toolbox Model for ResNet-50 Network support package. To install these support packages, use the MATLAB
The following line of code creates a folder in your current working folder on the host and copies all the relevant files into this folder. If you cannot generate files in this folder, before running this command, change your current working folder.
The support package uses an SSH connection over TCP/IP to execute commands while building and running the generated CUDA code on the Jetson platforms. Connect the target platform to the same network as the host computer or use an Ethernet crossover cable to connect the board directly to the host computer. For information on how to set up and configure your board, see NVIDIA documentation.
To communicate with the NVIDIA hardware, create a live hardware connection object by using the
jetson function. You must know the host name or IP address, user name, and password of the target board to create a live hardware connection object. For example, when connecting to the target board for the first time, create a live object for Jetson hardware by using the command:
hwobj = jetson('jetson-tx2-name','ubuntu','ubuntu');
During the hardware live object creation, the support package performs hardware and software checks, IO server installation, and gathers peripheral information on target. This information is displayed in the Command Window.
When there are multiple live connection objects for different targets, the code generator performs a remote build on the target board for which a recent live object was created. To choose a hardware board for performing a remote build, use the
setupCodegenContext() method of the respective live hardware object. If only one live connection object was created, you do not need to call this method.
To verify that the compilers and libraries necessary for running this example are set up correctly, use the
coder.checkGpuInstall (GPU Coder) function.
envCfg = coder.gpuEnvConfig('jetson'); envCfg.DeepLibTarget = 'cudnn'; envCfg.DeepCodegen = 1; envCfg.Quiet = 1; envCfg.HardwareObject = hwobj; coder.checkGpuInstall(envCfg);
To find the list of cameras connected to the target, use the
getCameraList function. To see the resolutions supported by the camera, use the
Available Resolutions column.
The Simulink model for classifying webcam images contains a
Predict block to predict the scores and labels for each class, an
NVIDIA Camera block to capture a live video stream, and an
NVIDIA Display block to display the results from the classification. Additional processing operations are implemented by using
MATLAB Function blocks.
Camera block captures a snapshot for each time-step of the algorithm. To select the webcam and set the resolution of each snapshot, double click on the
Camera block and set the
Image size as shown. The value of
Image size must be a resolution supported by the webcam.
rgb2Img function block converts the
B component outputs from the
Camera block to an RGB planar image. The resulting image is then provided as input to
resizeImg function block that resizes the image to the input image size of the Resnet-50 network. The input image size of the Resnet-50 network is 224-by-224-by-3.
predict block from Deep learning toolbox takes a single input image frame and runs prediction on the image by using the pretrained
resnet50 convolutional neural network. ResNet-50 is a DAG network trained on more than a million images from the
ImageNet database. The output contains the categorical scores of each class the image belongs to. Double click the
predict block to open block parameters and set the deep network as
resnet50 as shown.
calcFPS function block calculates the elapsed time between the execution of the
predict block and calculates the average frames of output per second (FPS) displayed on the target.
The output scores, average frames per second(FPS) of the output, and the input image are provided to a
insertResult function block to annotate the input image with the top five prediction labels, corresponding scores, and the average frames per second.
function [outR,outG,outB] = insertResult(inpImg, scores, fps)
[scores,index] = sort(scores,'descend');
%% Insert prediction with scores in the image
%Pad Image outImg = padarray(inpImg,[0,200],0,'pre');
% Load the Scores index5 = index(1:5); synsetData = coder.load('synsetWords.mat'); synsetData = synsetData.synsetArray;
% Insert Avg. FPS outImg = insertText(outImg, [10 80],['FPS : ' sprintf('%0.2f',fps) ], 'AnchorPoint','LeftBottom');
% Insert Top 5 predictions with corresponding scores for i=1:5 str = strtrim(synsetData(index5(i),:)); outImg = insertText(outImg, [10 50+30*(i+1)],[str ': ' sprintf('%0.2f',scores(i)*100) '%'], 'AnchorPoint','LeftBottom'); end
% Split the image into R,G,B components outR = outImg(:,:,1)'; outG = outImg(:,:,2)'; outB = outImg(:,:,3)';
The resulting output image is provided to the
NVIDIA Display block which opens a display window showing the output image on the target while running an application.
To generate CUDA code, the model must be enabled with the following settings:
The model in this example is preconfigured to generate CUDA code.
For more information on generating GPU Code in Simulink, see Code Generation from Simulink Models with GPU Coder (GPU Coder).
The Deep learning target library must be set as
'CuDNN' during code generation,
For more information, see Deep Learning in Simulink by Using MATLAB Function Block (GPU Coder) and Deep Learning in Simulink by Using Deep Neural Networks Library (GPU Coder).
The model configuration parameters provide options for build process and deployment.
1. Open the Configuration Parameters dialog box, Hardware Implementation pane. Select Hardware board as NVIDIA Jetson (or NVIDIA Drive).
2. On the Target hardware resources section, enter the Device Address, Username, and Password of your NVIDIA Jetson target board.
3. Set the Build action, Build directory, and Display as shown. The Display represents the display environment and allows output to be displayed to the device corresponding to the environment.
4. Click Apply and OK.
1. To generate CUDA code for the model, deploy the code to the target, and run the executable, open the Hardware tab on the Simulink Editor.
2. Select Build, Deploy & Start to generate and deploy the code on the hardware.
When the application starts on the target, an SDL window opens showing the classification results for each frame of the video stream captured by the webcam.
To stop the application on the target, use the
stopModel method of the hardware object.
1. This example uses cuDNN as the deep learning code target library. This can be changed to NVIDIA TensorRT by running the following command. This requires the NVIDIA TensorRT - high performance deep learning inference optimizer and run-time library must be installed on the Jetson platform.
2. This example sets the Hardware Implementation for NVIDIA Jetson. User can run this example with the NVIDIA Drive as target hardware.
3. The image resolution of the camera block can be changed based on the requirement.
close_system to close the model.
To remove the example files and return to the original folder, call the