Jetson-Ready ROS2 Object Detection

Small ROS2 object-detection pipeline for camera streams. It subscribes to camera images, runs a PyTorch detector, publishes vision_msgs/Detection2DArray bounding boxes, and publishes an annotated sensor_msgs/Image stream.

Jetson-oriented, not yet tested on Jetson.

Detection output

The screenshots are representative local sample outputs for the README. They are not Jetson benchmark claims. The street-scene photo is based on “Busy street during rush hour” by Victor Kibiwott, licensed under CC BY-SA 4.0, with annotation overlays added.

ROS2 deployment workflow

• A ROS2 Python node that consumes sensor_msgs/Image camera frames.
• Detector backend abstraction with torchvision Faster R-CNN by default and optional Ultralytics YOLO support.
• Bounding-box publishing through vision_msgs/Detection2DArray.
• Annotated image publishing for RViz, rqt_image_view, recording, or downstream visualization.
• A repeatable FPS/latency benchmark entry point for CPU, CUDA desktop GPU, or Jetson experiments.
• A Docker build path that keeps ROS2, OpenCV, PyTorch, and the package together.

Topics

Direction	Topic	Type
Subscribe	/camera/image_raw	sensor_msgs/Image
Publish	/detections	vision_msgs/Detection2DArray
Publish	/detections/annotated	sensor_msgs/Image
Publish	/detections/stats	diagnostic_msgs/KeyValue

Quick Start

Install ROS2 Humble or later with cv_bridge and vision_msgs, then install PyTorch and Torchvision for your machine.

colcon build --symlink-install
source install/setup.bash
ros2 launch jetson_ready_ros2_object_detection object_detection.launch.py \
  image_topic:=/camera/image_raw \
  backend:=torchvision \
  device:=auto

For YOLO:

pip install ultralytics
ros2 launch jetson_ready_ros2_object_detection object_detection.launch.py \
  backend:=yolo \
  model:=yolov8n.pt

View the annotated stream:

ros2 run rqt_image_view rqt_image_view /detections/annotated

Benchmark

Run repeated inference on a synthetic frame:

ros2 run jetson_ready_ros2_object_detection benchmark \
  --backend torchvision \
  --device auto \
  --warmup 5 \
  --iterations 50

Benchmark an actual image and save JSON:

ros2 run jetson_ready_ros2_object_detection benchmark \
  --backend yolo \
  --model yolov8n.pt \
  --image captures/frame.jpg \
  --device cuda \
  --json benchmark-results/yolov8n-cuda.json

The benchmark reports FPS, mean latency, median latency, p95 latency, frame shape, and the final detection count.

Docker

Build for a desktop CPU smoke test:

docker build -t ros2-object-detection:humble .

Run with host networking so the container can see ROS graph traffic:

docker run --rm -it --net=host --ipc=host ros2-object-detection:humble

For NVIDIA GPU runtime on a desktop or Jetson:

docker run --rm -it --net=host --ipc=host --gpus all ros2-object-detection:humble

Jetson builds should use PyTorch and Torchvision wheels that match the installed JetPack/L4T release:

docker build \
  --build-arg TORCH_INSTALL="<jetson-compatible torch torchvision install command>" \
  -t ros2-object-detection:jetson .

Configuration

Defaults live in config/detector.yaml. Common launch overrides:

Parameter	Default	Meaning
image_topic	/camera/image_raw	Input camera stream
backend	torchvision	torchvision or yolo
model	empty	Torchvision weights selector or YOLO model path
device	auto	auto, cpu, cuda, or cuda:0
score_threshold	0.5	Minimum confidence
max_detections	50	Per-frame detection cap
publish_annotated	true	Publish annotated image frames

Deployment caveats

• Jetson-oriented, not yet tested on Jetson.
• First run may download detector weights unless the model files are already cached.
• The default Torchvision model is easy to run but not optimized for Jetson latency.
• No TensorRT export path is included yet; adding ONNX/TensorRT engines is the natural next step.
• No object tracking, temporal smoothing, camera calibration, or 3D projection is included.
• QoS settings are intentionally simple and may need tuning for high-rate cameras.