CvGoodFeatureToTrack
lastmod 2015-05-05created 2015-05-051 edits
Corners
Points with large derivative values (points where pixel values change a lot) will be points near edges in the image. However, points on the same edge can all show similar characteristics, so they cannot be called unique points. So if large derivative values appear at the same time in mutually orthogonal directions, the point can be called very distinctive. Points like this, with large derivative values in orthogonal directions, are suitable for tracking, and such points are called corners.
Harris corner detection
Using the second derivatives of image brightness (the rate of change of the change in pixel values), a small window is set at each point and the autocorrelation matrix of the second-derivative image is computed within this window. If both eigenvalues of the autocorrelation matrix are large, the point is judged to be a corner. Shi and Tomasi: when the smaller of the two eigenvalues is larger than a certain threshold, this point is a good corner