1.安装ultralytics

1 2	pip install --upgrade pip pip install ultralytics

2.安装gpu版本的torch

安装好后，自带的torch是cpu版本的，建议是换成gpu
运行

1	pip uninstall -y torch torchvision

卸载torch

由于pytorch官网提供的linux torch架构是X86的，jetson nano为arrch64，所以不能从pytorch官网下载
在动物园可以找到wheel文件，但是在这我只找到了torch，没有torchvision

最后在这里找到了线索

1
2

pip install https://github.com/ultralytics/assets/releases/download/v0.0.0/torch-2.5.0a0+872d972e41.nv24.08-cp310-cp310-linux_aarch64.whl
pip install https://github.com/ultralytics/assets/releases/download/v0.0.0/torchvision-0.20.0a0+afc54f7-cp310-cp310-linux_aarch64.whl

这样就算是装完了

3.配置环境变量

因为在conda环境下，默认是读不到系统自带的cuda跟tensorrt路径的，我们需要自己配置一下。
注意在运行一些pip install时候，不要安装上cuda 跟tensorrt了，pip安装的是用不了的
我们只需要配置一下环境变量就行

#找路径
find /usr -name "tensorrt" -type d 2>/dev/null
ls -la /usr/local/cuda

#cuda环境变量

vim ~/.bashrc   #把下面的几行导入保存，
                #请把路径替换成自己的，直接复制粘贴可能不太好用

1
2
3

export LD_LIBRARY_PATH=/usr/lib/aarch64-linux-gnu:$LD_LIBRARY_PATH
export CUDA_HOME=/usr/local/cuda-12.6
export PATH=/usr/local/cuda-12.6/bin:$PATH

#tensorrt环境变量
mkdir -p $CONDA_PREFIX/etc/conda/activate.d
vim $CONDA_PREFIX/etc/conda/activate.d/tensorrt.sh
export PYTHONPATH=/usr/lib/python3.10/dist-packages:$PYTHONPATH
export LD_LIBRARY_PATH=/usr/lib/aarch64-linux-gnu:/usr/local/cuda/lib64:$LD_LIBRARY_PATH

#检测是否正常
python - << 'EOF'
import tensorrt as trt
import pycuda.driver as cuda
print(trt.__version__)
EOF

4.下载yolo模型,并进行格式转换

去这个地方就可以下到yolo的模型了

1	wget https://github.com/ultralytics/assets/releases/download/v0.0.0/yolov8n.pt

分别执行下面代码即可实现格式转换

1 2	#pt转onnx yolo export model=yolov8n.pt format=onnx device=cpu

#onnx转engine
/usr/src/tensorrt/bin/trtexec \
  --onnx=yolov8n.onnx \
  --saveEngine=yolov8n.engine \
  --fp16

5.实现模型推理

下面是例程

from ultralytics import YOLO
import cv2
import time

model = YOLO("yolov8n.engine", task="detect")

cap = cv2.VideoCapture("Delta.mp4")

while cap.isOpened():
    ret, frame = cap.read()
    if not ret:
        break

    t0 = time.time()

    results = model(frame)   # TensorRT engine 推理

    fps = 1.0 / (time.time() - t0)

    annotated = results[0].plot()
    cv2.putText(
        annotated,
        f"FPS: {int(fps)}",
        (30, 50),
        cv2.FONT_HERSHEY_SIMPLEX,
        1,
        (0, 255, 0),
        2
    )

    cv2.imshow("YOLOv8 TensorRT", annotated)
    if cv2.waitKey(1) & 0xFF == ord("q"):
        break

cap.release()
cv2.destroyAllWindows()

参考

Quick Start Guide: NVIDIA Jetson with Ultralytics YOLO11
Ultralytics YOLO Docs
Jetson Zoo
Installing PyTorch for Jetson Platform
nvidia pytorch download
Jetson nano部署YOLOV8并利用TensorRT加速推理实现行人检测追踪

yolo export model=yolov8n.pt format=engine device=0
yolo predict model=yolov8n.engine source=images/ device=0