Welcome to Jetson Orin NANO Developer Kit repository
01.DeepStream environment
DeepStream Environment Setup
DeepStream Environment Setup1. Pre-Installation Checks2. Downloading the DeepStream 9.1 Package3. Installing DeepStream 9.14. Verifying the Installation5. Running the Basic Video Detection Example6. USB and CSI Camera Detection6.1. USB Camera6.2. CSI Camera7. Cleaning Up the Installation Package
1. Pre-Installation Checks
First confirm that the system is already on the JetPack 7.2.1 baseline, and check the TensorRT dependencies. DeepStream 9.1 requires CUDA 13.2 and TensorRT 10.16.
cat /etc/os-releasedpkg-query -W -f='${Package} ${Version}\n' \ nvidia-l4t-core libnvinfer10 libnvonnxparsers10 libnvinfer-plugin10df -h /The target machine versions actually verified in this course:
xxxxxxxxxxUbuntu 24.04.4 LTSL4T R39.2.1CUDA 13.2TensorRT 10.16.2Before installation, reserve at least 5 GB of free space on the root partition; downloading advanced models and building TensorRT engines later will require more space.
2. Downloading the DeepStream 9.1 Package
The DeepStream 9.1 Jetson arm64 package is published on the official NVIDIA GitHub Release. Using aria2c with multiple connections for the download is recommended.
xxxxxxxxxxcd /tmparia2c -c -x 16 -s 16 -k 1M \ --check-integrity=true \ --dir=/tmp \ --out=deepstream-9.1_9.1.0-1_arm64.deb \ https://github.com/NVIDIA/DeepStream/releases/download/v9.1.0/deepstream-9.1_9.1.0-1_arm64.debAfter the download finishes, check the file:
xxxxxxxxxxls -lh /tmp/deepstream-9.1_9.1.0-1_arm64.debsha256sum /tmp/deepstream-9.1_9.1.0-1_arm64.deb3. Installing DeepStream 9.1
Use APT to install the local Debian package so that APT automatically resolves and installs the missing dependencies.
xxxxxxxxxxsudo apt-get install -y /tmp/deepstream-9.1_9.1.0-1_arm64.debWhen running deepstream-app for the first time, if it reports that libgstrtspserver-1.0.so.0 cannot be found, install the runtime dependency:
xxxxxxxxxxsudo apt-get install -y libgstrtspserver-1.0-0Do not copy commands from other environments that forcibly install, downgrade, or mix TensorRT versions. If a matching TensorRT 10.16 / CUDA 13.2 already exists before installation, simply install DeepStream directly.
The installation directory is:
xxxxxxxxxx/opt/nvidia/deepstream/deepstream-9.1/4. Verifying the Installation
Check the DeepStream, CUDA, and TensorRT versions:
xxxxxxxxxxdeepstream-app --version-allThe expected output includes:
xxxxxxxxxxdeepstream-app version 9.1.0DeepStreamSDK 9.1.0CUDA Driver Version: 13.2CUDA Runtime Version: 13.2TensorRT Version: 10.16
Confirm the package status and the core GStreamer plugins:
xdpkg-query -W -f='${Status} ${Version}\n' deepstream-9.1
gst-inspect-1.0 nvinfer >/dev/null && echo "nvinfer OK"gst-inspect-1.0 nvstreammux >/dev/null && echo "nvstreammux OK"gst-inspect-1.0 nvv4l2decoder >/dev/null && echo "Hardware decoder plugin OK"5. Running the Basic Video Detection Example
This example uses the traffic video bundled with DeepStream and TrafficCamNet; the terminal continuously outputs the number of vehicles and persons in each frame.
xxxxxxxxxxcd /opt/nvidia/deepstream/deepstream-9.1/sources/apps/sample_apps/deepstream-test1sudo deepstream-test1-app dstest1_config.ymlThe full path of the test video is:
xxxxxxxxxx/opt/nvidia/deepstream/deepstream-9.1/samples/streams/sample_720p.h264Signs of success:
xxxxxxxxxxFrame Number = 0 Number of objects = ... Vehicle Count = ... Person Count = ......End of streamReturned, stopping playbackDeleting pipeline
The following messages on the first run are normal:
xxxxxxxxxxDeserialize engine failedTrying to create engine from model filesThis means the corresponding TensorRT engine does not yet exist on the device, and DeepStream is building the engine from the ONNX model. The first build is slower.
If failed to serialize cuda engine to file appears, it usually means the regular user does not have write permission for /opt/nvidia/deepstream/.../samples/models; this inference still runs normally, but subsequent launches may rebuild the engine.
6. USB and CSI Camera Detection
Run this in a graphical desktop, VNC, or RDP session. Pure SSH has no display environment and cannot verify the nv3d-sink graphical output.
6.1. USB Camera
You can first confirm the camera node:
xxxxxxxxxxls -l /dev/video*Enter the following commands to preview:
xxxxxxxxxxcd /opt/nvidia/deepstream/deepstream-9.1/samples/configs/deepstream-appsudo deepstream-app -c source1_usb_dec_infer_resnet.txt6.2. CSI Camera
xxxxxxxxxxcd /opt/nvidia/deepstream/deepstream-9.1/samples/configs/deepstream-appsudo deepstream-app -c source1_csi_dec_infer_resnet.txtWhen the CSI camera succeeds, the terminal lists the sensor modes supported by Argus and continuously outputs FPS. If Argus recognizes the sensor but the popup window goes black and exits, check the CSI camera module, flex cable, and connector first; in this actual test, replacing the CSI camera made it run normally.
7. Cleaning Up the Installation Package
After installation and verification are complete, you can delete the temporary installation package and download log in /tmp to avoid taking up space:
xxxxxxxxxxrm -f /tmp/deepstream-9.1_9.1.0-1_arm64.deb \ /tmp/deepstream-9.1_9.1.0-1_arm64.deb.aria2 \ /tmp/deepstream-9.1-download.logDo not run sudo apt autoremove unless you have confirmed that the listed automatically installed packages are no longer used by other courses or projects.