AVL.DetectCorners_CornerResponse(AvlNet.Image, AvlNet.Region, AvlNet.CornerResponseMethod, int, float?, AvlNet.Point2D[], AvlNet.Image)

The operation detects corners using either Harris or Kanade-Tomasi corner response method, depending on inCornerResponseMethod. For every square window of size inKernelSize a convolution matrix is computed: \[M = \left(\begin{array}{ccc} \sum g_r^2 & \sum g_c g_r \\ \sum g_c g_r & \sum g_c^2 \end{array} \right) \] where the summation is performed over the whole window and \(g_r, g_c\) denote horizontal and vertical gradient respectively at the point.
Harris' corner response is computed the following way: \[H = \mathrm{det}(M) - k \mathrm{tr}^2(M) \] where k is a constant set to 0.01.
Kanade-Tomasi corner response is given by: \[KT = \mathrm{min}(\lambda_1, \lambda_2)\] where \(\lambda_1, \lambda_2\) are eigenvalues of the convolution matrix.
Values H or KT give corner response image.

Then a few steps are performed in order to extract corner points. First, the normalized corner response function is thresholded with inThreshold with a small hysteresis, then the remained points are split into connected regions (blobs) and the center of each blob is determined.

DetectCorners_CornerResponse with inThreshold=50, methods Harris and Kanade-Tomasi respectively.

Both methods give similar results and are quite fast. However, they tend to be less accurate than DetectCorners_Foerstner.