algorithmmatlabcomputer-visionfeature-detectioncorner-detection

# Implementing a Harris corner detector

I am implementing a Harris corner detector for educational purposes but I'm stuck at the harris response part. Basically, what I am doing, is:

1. Compute image intensity gradients in x- and y-direction
2. Blur output of (1)
3. Compute Harris response over output of (2)
4. Suppress non-maximas in output of (3) in a 3x3-neighborhood and threshold output

1 and 2 seem to work fine; however, I get very small values as the Harris response, and no point does reach the threshold. Input is a standard outdoor photography.

``````[...]
g = fspecial('gaussian');
Ix = imfilter(Ix, g);
Iy = imfilter(Iy, g);
H = harrisResponse(Ix, Iy);
[...]

function K = harrisResponse(Ix, Iy)
max = 0;
[sy, sx] = size(Ix);
K = zeros(sy, sx);
for i = 1:sx,
for j = 1:sy,
H = [Ix(j,i) * Ix(j,i), Ix(j,i) * Iy(j,i)
Ix(j,i) * Iy(j,i), Iy(j,i) * Iy(j,i)];
K(j,i) = det(H) / trace(H);
if K(j,i) > max,
max = K(j,i);
end
end
end
max
end
``````

For the sample picture, max ends up being 6.4163e-018 which seems far too low.

Solution

• A corner in Harris corner detection is defined as "the highest value pixel in a region" (usually `3X3` or `5x5`) so your comment about no point reaching a "threshold" seems strange to me. Just collect all pixels that have a higher value than all other pixels in the `5x5` neighborhood around them.

Apart from that: I'm not 100% sure, but I think you should have:

`K(j,i) = det(H) - lambda*(trace(H)^2)` Where lambda is a positive constant that works in your case (and Harris suggested value is 0.04).

In general the only sensible moment to filter your input is before this point:

`[Ix, Iy] = intensityGradients(img);`

Filtering `Ix2`, `Iy2` and `Ixy` doesn't make much sense to me.

Further, I think your sample code is wrong here (does function `harrisResponse` have two or three input variables?):

``````H = harrisResponse(Ix2, Ixy, Iy2);
[...]

function K = harrisResponse(Ix, Iy)
``````