Corner and Centroid Detection

Preface

This tutorial is applicable to firmware version: CanMV_K230_YAHBOOM_micropython_v1.8-13. Download from Jianan official website https://download.kendryte.com/developer/releases/canmv_k230_micropython/daily_build/

Example Results

Run the example code for this section 04.camera_corners.py.

This section performs grayscale conversion and OTSU binarization on the real-time camera feed, extracts external contours, calculates contour centroids using image moments, and obtains contour corners using polygon approximation. The centroid is marked with a blue cross, and corners are marked with green solid circles.

Principle Explanation

What are Image Moments and Centroids

Image moments are a set of features obtained by weighted summation of contour coordinates and their distributions. The zeroth-order moment m00 corresponds to the contour area, and the first-order moments m10 and m01 can be used to calculate the centroid:

cx = m10 / m00
cy = m01 / m00

When m00 is 0, division is not possible, so the code must first check if m00 > 0.

Why Polygon Approximation Can Find Corners

approxPolyDP() uses the Douglas-Peucker algorithm to simplify contours, retaining only the points that have significant impact on the overall shape. A smaller error threshold preserves more points; a larger threshold simplifies the contour more. In this example, the error is set to 2% of the contour perimeter.

Note that this example obtains "contour approximation vertices", not texture corners in the sense of goodFeaturesToTrack() or Harris algorithm.

Code Overview

Initialize Camera and Display Module

xxxxxxxxxx
W, H = 640, 480
sensor = Sensor(width=1280, height=960, fps=90)
sensor.reset()
sensor.set_framesize(width=W, height=H)
sensor.set_pixformat(Sensor.RGB888)
Display.init(Display.ST7701, width=640, height=480, to_ide=True)
sensor.run()
time.sleep(0.5)

Set Contour Area Threshold

xxxxxxxxxx
MIN_A = 200

Contours with area less than 200 pixels are ignored directly to reduce corners from binary noise.

Grayscale, Binarization, and Contour Extraction

xxxxxxxxxx
gray = cv2.cvtColor(img_np, cv2.COLOR_BGR2GRAY)
ret, binary = cv2.threshold(gray, 0, 255, cv2.THRESH_OTSU)
contours, _ = cv2.findContours(
    binary,
    cv2.RETR_EXTERNAL,
    cv2.CHAIN_APPROX_SIMPLE
)

OTSU auto-threshold is suitable for scenes where foreground and background have significantly different grayscale distributions. RETR_EXTERNAL only takes external contours, and internal holes will not be processed as independent contours.

Calculate and Draw Centroids

xxxxxxxxxx
moments = cv2.moments(contour)
if moments["m00"] > 0:
    cx = int(moments["m10"] / moments["m00"])
    cy = int(moments["m01"] / moments["m00"])
    cv2.drawMarker(
        img_np, (cx, cy),
        (255, 0, 0),
        cv2.MARKER_CROSS, 10, 2
    )

Centroid coordinates are converted to integers for drawing. The color tuple is passed according to the current image channel order.

Extract and Draw Corners

​x
perimeter = cv2.arcLength(contour, True)
approx = cv2.approxPolyDP(contour, 0.02 * perimeter, True)

for point in approx:
    x = int(point[0][0])
    y = int(point[0][1])
    cv2.circle(img_np, (x, y), 5, (0, 255, 0), -1)

cv2.drawContours(img_np, [contour], -1, (255, 255, 0), 1)

arcLength(contour, True) treats the contour as a closed curve. The source code processes up to 50 valid contours at most to prevent excessive drawing from complex backgrounds.

Display, Statistics, and Resource Release

xxxxxxxxxx
Display.show_image(img)

if time.ticks_diff(time.ticks_ms(), ft) >= 1000:
    print("corners:", corner_count, "| fps:", fc)

finally:
    sensor.stop()
    Display.deinit()
    os.exitpoint(os.EXITPOINT_ENABLE_SLEEP)
    time.sleep_ms(100)

Parameter Adjustment Notes

• MIN_A: Decrease when small target corners are filtered out, increase when there are too many small contours in the background.

• OTSU Foreground Polarity: If the target is divided into black, background external contours may be detected. Use THRESH_BINARY_INV or a fixed threshold.

• Approximation Error 0.02: Increase when there are too many corners, decrease when corners are being merged.

• Contour Upper Limit: The source code limit is 50. Actual projects should use ROI and area range to reduce candidate contours in advance.

• Texture Corners: When detecting chessboard patterns, textures, or corners inside objects, use goodFeaturesToTrack() instead of treating contour vertices as corners for all types.