YOLOv8 + TensorRT ROS Inference (Dockerized)

This repository contains a minimal and production-ready Docker environment to run YOLOv8 inference in real-time using TensorRT, integrated with ROS (Noetic) for robotics workflows. Built for Jetson or remote GPU inference, with CPU fallback.

---

🚀 Docker Image on Docker Hub

📦 DockerHub: pranaypalem/yolov8_tensorrt

docker pull pranaypalem/yolov8_tensorrt:3.1.3

---

🔧 Features

• ✅ ROS Noetic (minimal setup)
• ✅ Inference via TensorRT engine (or ONNXRuntime as fallback)
• ✅ Subscribes to image topic from ROS
• ✅ Publishes annotated image with YOLO detections
• ✅ GPU + CPU fallback support
• ✅ Caches .engine file for fast reuse

---

🛠️ Requirements

• Docker with NVIDIA GPU support
• ROS master running (Jetson or another device)
• YOLOv8 best.pt model file
• TensorRT tar.gz installer must be manually downloaded:

  Download TensorRT from: NVIDIA Developer TensorRT

  Place TensorRT-8.6.1.6.Linux.x86_64-gnu.cuda-11.8.tar.gz into the project directory before building or running.

---

🧪 How to Run

1. Launch the Docker container:

docker run -it --gpus all --net=host \
  -e ROS_MASTER_URI=http://172.20.10.4:11311 \
  -e ROS_IP=172.20.10.6 \
  pranaypalem/yolov8_tensorrt:3.1.3

2. Convert the YOLOv8 model to TensorRT Engine:

Inside the container:

python3 convert_to_engine.py

This will convert your .pt model to an optimized .engine file.

3. Run the YOLO ROS inference node:

python3 yolo_ros_inference.py

This script subscribes to a ROS image topic, runs YOLOv8 inference using TensorRT, and publishes the annotated image back.

---

📁 Folder Structure

.
├── Dockerfile
├── bestsynthetic.pt
├── coke_can_depth_processor.py
├── convert_to_engine.py
├── test_video.avi
├── video_publisher.py
├── yolo_ros_inference.py
├── TensorRT-8.6.1.6.Linux.x86_64-gnu.cuda-11.8.tar.gz   # Must be manually downloaded
└── README.md

---

📂 Environment Variables

Variable	Description
ROS_MASTER_URI	IP address of the ROS master node
ROS_IP	Local machine IP to advertise to ROS

---

🧠 Internals

• Converts best.pt → best.onnx → best.engine
• Runs YOLOv8 inference using TensorRT (or ONNXRuntime as fallback)
• Publishes annotated images back to ROS topics

---

🧑‍💻 Author

Pranay Palem
Optimized for real-time robotics, computer vision pipelines, and GPU inference workflows.

---

📜 License

MIT License — use freely and responsibly.