Skip to main content

Model Prediction on Jetson

This section explains how to perform model prediction (inference) on NVIDIA Jetson using trained YOLO models in PyTorch, ONNX, or TensorRT formats.

1. Best Performance Mode​

Before running inference, configure Jetson for maximum performance.

1.1 Enable MAX Power Mode​

sudo nvpmodel -m 2

1.2 Enable Jetson Clocks​

sudo jetson_clocks

2. Model Prediction​

2.1 CLI Usage​

note

The YOLO CLI currently supports USB cameras only. For CSI cameras, use Python-based inference.

Run prediction with a TensorRT engine model:

yolo predict model=best.engine source=0 save=False show
  • source=0: USB camera index\
  • For multiple cameras, increment the index accordingly

Output videos are saved to:

/home/jetson/ultralytics/ultralytics/output/

CLI Prediction Result

2.2 Python Usage​

Python-based inference supports both USB cameras and CSI cameras.

2.2.1 USB Camera Prediction​

cd /home/jetson/ultralytics/ultralytics/yahboom_demo

Run USB Camera Script​

python3 06.orange_camera_usb.py

Press q to exit the preview window.

Output Preview​

/home/jetson/ultralytics/ultralytics/output/

USB Camera Result

Sample Code (USB Camera)​

import cv2
from ultralytics import YOLO

model = YOLO(
    "/home/jetson/ultralytics/ultralytics/data/yahboom_data/orange_data/"
    "runs/detect/train/weights/best.engine"
)

cap = cv2.VideoCapture(0)

frame_width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
frame_height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
fps = int(cap.get(cv2.CAP_PROP_FPS))

out = cv2.VideoWriter(
    "/home/jetson/ultralytics/ultralytics/output/orange_usb.mp4",
    cv2.VideoWriter_fourcc(*"mp4v"),
    fps,
    (frame_width, frame_height),
)

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

    results = model(frame)
    annotated = results[0].plot()
    out.write(annotated)

    cv2.imshow("YOLO Inference", cv2.resize(annotated, (640, 480)))
    if cv2.waitKey(1) & 0xFF == ord("q"):
        break

cap.release()
out.release()
cv2.destroyAllWindows()

2.2.2 CSI Camera Prediction​

cd /home/jetson/ultralytics/ultralytics/yahboom_demo

Run CSI Camera Script​

python3 06.orange_camera_csi.py

Press q to exit the preview window.

Sample Code (CSI Camera)​

import cv2
from ultralytics import YOLO
from jetcam.csi_camera import CSICamera

model = YOLO(
    "/home/jetson/ultralytics/ultralytics/data/yahboom_data/orange_data/"
    "runs/detect/train/weights/best.engine"
)

camera = CSICamera(width=640, height=480)

while True:
    frame = camera.read()
    results = model(frame)
    annotated = results[0].plot()

    cv2.imshow("YOLO CSI Inference", annotated)
    if cv2.waitKey(1) & 0xFF == ord("q"):
        break

cv2.destroyAllWindows()

CSI Camera Result

Summary​

  • Supports inference using .pt, .onnx, and .engine models
  • CLI inference is limited to USB cameras
  • Python inference supports USB and CSI cameras
  • TensorRT (.engine) provides the best performance

Maintained by HemiHex for Jetson-based advanced vision workflows.