Yolo11/26 C++ TRT Object Detection on Jetson Orin Nano

Robotics & Edge Computing Jetson Systems Jetson Orin Nano
,
1

I’m looking for a lightweight Yolo11/26 C++ TRT Deepstream example for object detection for 1 camera to build from source (no dockers). Please help. I have an official jetson orin nano devkit with: JetPack 6.2.1, L4T:3 6.4.4, Cuda: 12.6.68, TensorRT: 10.3.0.30, setup details: system_info.txt (1.6 KB). A few projects on github are available for windows and jetson devices, but they don’t compile on my system.

My project requires very low latency, so I need a C++ foundation, not python. I did get Qengineering’s Yolo8-TensorRT-C++ project to build and run, but it’s limited to Yolo8.

A C++ example that is lightweight, works with preinstalled TRT, Cuda, and OpenCV, and will port smoothly to upcoming JetPacks, is highly desirable.

6

Please refer to this repository.It is based on deepstream-app.

If the performance of this project does not meet your needs, you can try int8 quantization, setting the nvinfer interval parameter, and optimizing the decoding parameters to match nvstreammux/nvinfer, etc.

7

The github page your referenced only has 1 c++ example with the outdated yolov7 model. I’m looking for a c++ example for Yolov11 or Yolov26, specifically for the orin nano devkit with JetPack 6.2.1 per the original request. I notice the C++ code varies considerably between yolo models and especially TensorRT versions.

Please move this thread back to the Jetson Orin Nano forum. It’s more likely to receive useful responses for me and the user community.

9

Moving back to Jetson Nano forum.

11

The link I provided above already includes YOLOv11s; for YOLOv26, only a few minor modifications are needed.

12

@junshengy, I saw the yolov11 link, but that is just a script to run deepstream.

I’m looking for a yolov11 C++ example that integrates deepstream, TRT, and yolov11 for the orin nano devkit. My goal is to add to that C++ program to stream inference results to another C++ program as part of my robotics project. The project requires the fastest inference speed possible, which is why I want this in C++, not python.

14

Hi,

Please try the example shared in the Ultralytics tutorial:

You can export the model to a TensorRT engine so it uses TensorRT for backend inference.

$ yolo export model=yolo11n.pt format=engine
$ yolo predict model=yolo11n.engine source='https://ultralytics.com/images/bus.jpg'

Thanks

15

I can already convert yolov11 *.pt models of various sizes (n, s, m, l, x) to TRT *.engine models with the ultralytics ecosystem and run inferencing, but in python. To be quite clear, I need to run inferencing in C++ similar to Nvidia’s Hello AI World project (which is sadly outdated).

I’m looking for a C++ example that I compile natively on my orin nano devkit to run inferencing with Yolo11 or Yolo26 models with CUDA and TRT, for the lowest possible latency necessary for my robotics project. Sadly the Hello AI World is outdated and does not work beyond Jetpack 6.0, and uses outdated networks (caffe, onnx, TAO). An updated C++ version with Yolo11 or 26 would be great for the user community interested in low latency capabilities.

16

Hi,

You can find some examples on their GitHub:

Thanks.