jalywang123 / yolo-tensorrt-jetson

Modified and customized version of "Jetson Nano: Deep Learning Inference Benchmarks Instructions"

Video YOLO with TensorRT on Jetson Nano

Modified and customized version of Jetson Nano: Deep Learning Inference Benchmarks Instructions.

Run real-time object detections on Jetson Nano with TensorRT optimized YOLO network.

Performance

Network	Framerate
YOLOv3	2 to 5
YOLOv3-tiny	24

YOLOv3-tiny is way faster but yields poor detection results.

Setup

Install prerequisites and fetch weights:

chmod +x prebuild.sh
./prebuild.sh

Build & Run

cd src
mkdir build && cd build
cmake -D CMAKE_BUILD_TYPE=Release ..
make
cd ../../
./src/build/trt-yolo-app --flagfile=/path/to/config-file.txt

# e.g.
./src/build/trt-yolo-app --flagfile=config/yolov3-tiny.txt

Refer to sample config files yolov2.txt, yolov2-tiny.txt, yolov3.txt and yolov3-tiny.txt in config/ directory.

GStreamer Reference

This project uses GStreamer to utilize Nvidia's hardware acceleration on video capture, encoding and decoding. You should change some code in src/main.cpp to reflect the camera setup you are using for Jetson Nano.

For example, to test I have the following run on Jetson Nano:

gst-launch-1.0 nvarguscamerasrc ! \
    'video/x-raw(memory:NVMM), width=(int)1920, height=(int)1080, format=(string)NV12, framerate=(fraction)30/1' ! \
    nvvidconv flip-method=0 ! 'video/x-raw, format=(string)BGRx' ! \
    videoconvert ! 'video/x-raw, format=(string)BGR' ! \
    videoconvert ! 'video/x-raw, format=(string)RGB' ! \
    videoconvert ! nvvidconv ! 'video/x-raw(memory:NVMM), format=(string)NV12' ! \
    nvv4l2h265enc insert-sps-pps=true ! 'video/x-h265, stream-format=(string)byte-stream' ! \
    queue ! h265parse ! queue ! \
    rtph265pay ! queue ! \
    udpsink host=192.168.1.194 port=1234

The command above read frames from the camera and then send the stream to 192.168.1.194 which is my desktop address in LAN.

From my desktop, run:

gst-launch-1.0 udpsrc port=1234 ! \
    application/x-rtp,encoding-name=H265 ! queue ! \
    rtph265depay ! queue ! avdec_h265 ! queue ! autovideosink

There you can see the streaming video on your desktop, which is being captured on Jetson Nano.

This means the GStreamer pipeline is valid, so you could use those commands in OpenCV VideoCapture's src and VideoWriter's des. See src/main.cpp for details.

For more on GStreamer and nvarguscamerasrc, see Accelerated GStreamer User Guide.