Developed with Giuseppe Marino
Given an image where
represents the coordinates of a pixel, it's possible to detect if the local neighbourhood of
contains a corner by using the Harris corner detector.
The standard implementation works by using the autocorrelation matrix
where and
are the partial derivatives of
. The determinant and the trace of this matrix are high respectively when there is a corner and when there is an edge or corner. Thus, we can define the matrix
such that
when there is a corner and
when there is an edge as
with a normalisation factor . We can then define a corner as a point
in
such that
is a local maximum and that
.
This implementation is only meant to be used to study the topic. It has a fixed value for k = 0.04 and takes t and n as parameters, where n represents the size of the window in which to perform the operations (e.g. if p is the center of the window and n = 2, then the window will go from p-2 to p+2 in both directions, so the size will be 5*5).
[H, keypoints, result] = harris(img, n, t);
figure; imshow(H, []);
figure; imshow(result);The function takes a grayscale img and returns the matrix H of size [h, w], the list of keypoints which consists of two columns (representing the indexes i, j of each corner pixel), and the image result obtained by overlaying the keypoints (represented as red pixels) over the input image.
 Original |
 Harris transform with n=3 and t=0.01 |
 Original overlayed with keypoints (result) |
|---|
The algorithm used comes from the section 2.1 in the paper Improving Harris corner selection strategy. The LaTeX syntax in this file is rendered with github-latex-markdown.