Run Yolo11n on Rock with RKNN

Refer to Radxa rock 5C documentation for futher details.

Introduction

This guid is using Radxa 4k camera camera to run Yolo11n on Rock 5C in realtime ( above 25FPS). This is very important for any visual based robotic need realtime decision making based on what it see.

Prerequirement

Hardware:

• Rock chip board like Rock 5C

Wiring guide

• Rock VIN → 5V
• Rock Ground → Ground

Run Code

• Create Virtual Environment

• Install ultralytics:

pip3 install -U ultralytics

• Test Basic prediction:

yolo predict model='./yolo11n_rknn_model' source='https://ultralytics.com/images/bus.jpg'

• Create python file:

nano run.py

• Add code below:

run.py

import cv2
import time
from ultralytics import YOLO
import numpy as np

# Load the RKNN model (must be in the same directory as the script)
model = YOLO("yolo11n_3588_rknn_model")

# Open the Radxa 4K MIPI-CSI camera (usually /dev/video0)
cap = cv2.VideoCapture("/dev/video11", cv2.CAP_V4L2)
cap.set(cv2.CAP_PROP_FRAME_WIDTH, 640)
cap.set(cv2.CAP_PROP_FRAME_HEIGHT, 640)

if not cap.isOpened():
    raise RuntimeError("❌ Failed to open Radxa 4K camera at /dev/video0")

# FPS monitor
frame_count = 0
start_time = time.time()

print("Running YOLOv11n Inference (press 'Q' to quit)...")

try:
    while True:
        ret, frame = cap.read()
        if not ret:
            print("⚠️ Camera frame capture failed.")
            continue

        # Resize to 640x640 and convert to RGB
        resized = cv2.resize(frame, (640, 640))

        # rgb = cv2.cvtColor(resized, cv2.COLOR_BGR2RGB)

        # Inference using RKNN-accelerated YOLOv11
        results = model.predict(source=resized, imgsz=640, stream=False)

        # Annotate and display
        annotated_frame = results[0].plot()
        cv2.imshow("YOLOv11n Detection", annotated_frame)

        if cv2.waitKey(1) & 0xFF == ord('q'):
            break

        # FPS Calculation
        frame_count += 1
        elapsed = time.time() - start_time
        if elapsed >= 1.0:
            fps = frame_count / elapsed
            print(f"🔄 FPS: {fps:.2f}")
            frame_count = 0
            start_time = time.time()

except KeyboardInterrupt:
    print("Inference stopped by user.")

finally:
    cap.release()
    cv2.destroyAllWindows()

Results