What are some easy to detect reference mark shapes and corresponding detection methods?

Question

Question Details:

There are a variety of image processing algorithms useful for detecting reference marks (contour tracing, FAST, MSER, Haar Feature-based Cascade Classifier, or even just Harris Edge detector, etc).

There are a variety of simple geometric shapes one might pick for use as reference marks (circle, square, X, etc).

I am wondering if someone has some experience with what combinations work best? Some additional criteria I have:

• using an algorithm implemented in openCV would be great

• Background invariance: detection algorithm should be robust against random background features(folds in clothing, random sharp corners, random colors present in background, speckle noise)

• Distance invariance: detection algorithm should be stable with respect to scale. (i.e. is should work whether the reference mark takes up a 100x100 pixel square in the image or 30x30 pixel square)

• Angle invariance: detection algorithm should not fail if the reference mark is viewed at an angle

• I am using a 640x480 webcam as a source of the images, so image processing at better than 5-10 fps would be nice

Additional Info:

My goal is to print out a fixed number black shapes on white paper. Place them in arbitrary positions in the scene that I am imaging with a webcam, and try to detect the center coordinates of each of the black shapes using a c++ program w/ openCV.

I am wondering if there is any consensus about what type of shape is the easiest to detect and how to do it.

Approaches that I have tried:

Detecting concentric circles using contour tracing: 1) Find the contours that are n times nested 2) compute the center point. The problem is that it is hard to resolve the rings at far distances (>5 feet) and angles.

Detecting cross shapes using contour tracing: 1) find contours, 2) find convex hull of each contour, 3) find number of convexity defects 4) compute the center if convexity defects == 4. Alternatively I have tried finding the depth of each convexity defect and comparing it to the width of the bounding rectangle around the contour to determine if it is a "real" convexity defect or just due to noise. This method works but sometimes certain background features, like folds in clothing, can make extra objects detected.

Answer 1

These marks (in the electronics manufacturing industry) are usually called fiducials or fiducial markers.

This paper from Accutron suggests:

Though I have no reference (or even measure) of the "optimality" of the disk / circle.

Answer 2

Even though this question is rather old, I'd like to give one certain recent development, which is related. Recently a fiducial tag detection system, termed X-tag is published:

X-Tag: A Fiducial Tag for Flexible and Accurate Bundle Adjustment, Tolga Birdal, Ievgeniia Dobryden and Slobodan Ilic, 3DV 2016.

There, the most suitable image features for detection is found to be ellipses and the paper presents extensive experiments on the angle and distance, comparing different solutions. The detection is based on a smart indexing scheme and is done completely at the image space, so that no marker coding is necessary. Therefore, the whole scheme is applicable to image-like scenarios. The accuracy is also found to be higher than the state of the art.

Plus, you have the luxury to generate random marks and just place them on a white background. Because the method involves hypotheses rejection, no outliers can get in the way.