Object Detection

---

1. Overview

YOLO11 is Ultralytics' next-gen object detection model with exceptional speed and accuracy. Local deployment on NVIDIA Jetson devices (Orin Nano/NX/AGX) enables efficient, low-latency AI inference.

Guide covers:

• Environment setup and JetPack installation
• Quick YOLO11 run via Docker
• Local YOLO11 installation
• TensorRT acceleration
• DLA acceleration and benchmarks

YOLO11 runs with ultra performance on Jetson Orin Nano, ideal for edge AI.

---

2. Environment Setup

Hardware Support

Device	JetPack Version	AI Performance
Jetson Nano	JetPack 4.6.x	472 GFLOPS
Jetson Xavier NX	JetPack 5.1.x	21 TOPS
Jetson Orin NX 16GB	JetPack 6.x	100 TOPS
Jetson Orin Nano Super	JetPack 6.x	67 TOPS

Recommended JetPack ≥5.1, enable max performance:

sudo nvpmodel -m 0
sudo jetson_clocks

---

3. Quick Start with Docker (Recommended)

Fastest method: Use pre-built Ultralytics image.

sudo docker pull ultralytics/ultralytics:latest-jetson-jetpack6
sudo docker run -it --ipc=host --runtime=nvidia ultralytics/ultralytics:latest-jetson-jetpack6

Includes YOLO11, PyTorch, Torchvision, TensorRT.

---

4. Local Installation (Optional)

For custom environments.

Step 1: Python Environment

sudo apt update
sudo apt install python3-pip -y
pip install -U pip

Step 2: Install YOLO11

pip install ultralytics[export]

Step 3: Install PyTorch and Torchvision

For JetPack 6.1 + Python 3.10:

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

Install cuSPARSELt for torch 2.5.0:

wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/arm64/cuda-keyring_1.1-1_all.deb
sudo dpkg -i cuda-keyring_1.1-1_all.deb
sudo apt-get update
sudo apt-get -y install libcusparselt0 libcusparselt-dev

---

Verify torch and GPU:

python3 -c "import torch; print(torch.__version__)" # 2.5.0a0+872d972e41.nv24.08
python3 -c "import torch; print(torch.cuda.is_available())" # True

Step 4: Install onnxruntime-gpu

pip install https://github.com/ultralytics/assets/releases/download/v0.0.0/onnxruntime_gpu-1.20.0-cp310-cp310-linux_aarch64.whl

5. TensorRT Acceleration

Ultralytics supports exporting models to TensorRT (.engine) for performance boost.

Python Example

from ultralytics import YOLO

model = YOLO("yolo11n.pt")
model.export(format="engine")  # Generates yolo11n.engine

trt_model = YOLO("yolo11n.engine")
results = trt_model("https://ultralytics.com/images/bus.jpg")

CLI Example

# Export to TensorRT
yolo export model=yolo11n.pt format=engine
# Run inference
yolo predict model=yolo11n.engine source='https://ultralytics.com/images/bus.jpg'

---

6. DLA (Deep Learning Accelerator)

Jetson devices feature DLA for lower power and higher concurrency.

Python Example

model.export(format="engine", device="dla:0", half=True)

CLI Example

# Export with DLA
yolo export model=yolo11n.pt format=engine device="dla:0" half=True
# Run on DLA
yolo predict model=yolo11n.engine source='https://ultralytics.com/images/bus.jpg'

Some layers may fall back to GPU if unsupported by DLA.

7. Object Detection Example

import cv2
import time
from ultralytics import YOLO

# Load TensorRT model
model = YOLO("yolo11n.engine") 

# Open USB camera
cap = cv2.VideoCapture(0)
cap.set(cv2.CAP_PROP_FRAME_WIDTH, 640)
cap.set(cv2.CAP_PROP_FRAME_HEIGHT, 480)

# FPS calculation
fps = 0.0
frame_count = 0
start_time = time.time()

print("📸 Real-time detection started. Press 'q' to quit.")
while True:
    ret, frame = cap.read()
    if not ret:
        break
    
    t0 = time.time()
    results = model(frame)
    annotated = results[0].plot()
    
    frame_count += 1
    t1 = time.time()
    fps = 1. / (t1 - t0)
    
    cv2.putText(annotated, f"FPS: {fps:.2f}", (10, 30),
                cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 0), 2)
    
    cv2.imshow("YOLO11 - TensorRT Real-time Detection", annotated)
    if cv2.waitKey(1) & 0xFF == ord("q"):
        break

cap.release()
cv2.destroyAllWindows()

---

7. Benchmark Comparison

Model Format	Orin Nano (ms)	mAP50-95	Orin NX (ms)
PyTorch	21.3	0.6176	19.5
TorchScript	13.4	0.6100	13.03
TensorRT (FP16)	4.91	0.6096	4.85
TensorRT (INT8)	3.91	0.3180	4.37

💡 TensorRT INT8 fastest, FP16 better accuracy.

---

8. Performance Tuning

Optimization	Recommendation
Power Mode	sudo nvpmodel -m 0
CPU/GPU Freq	sudo jetson_clocks
Monitoring	sudo pip install jetson-stats → jtop
Memory	Proper swap allocation

---

9. Common Issues

Issue	Solution
Can't import PyTorch	Use correct Jetson .whl
TensorRT slower than expected	Enable jetson_clocks + FP16
Can't pull Docker image	Ensure --runtime=nvidia
No tensorrt in venv	Copy from host: cp -r /usr/lib/python3.10/dist-packages/tensorrt your_venv/lib/python3.10/site-packages/

---

Appendix

• Ultralytics YOLO11 Docs
• Jetson Forum
• PyTorch for Jetson