An explanation of the Shi-Tomasi corner detection algorithm, covering why corners are the best features to track in computer vision applications like drones, AR, autonomous vehicles, and panorama stitching. The post walks through the autocorrelation function, Taylor expansion approximation, the second moment matrix (structural tensor), eigenvalue analysis for classifying smooth regions vs. edges vs. corners, and the full detection pipeline including gradient computation, Gaussian weighting, thresholding, and non-maximum suppression. It also compares Shi-Tomasi's smallest eigenvalue criterion with Harris corner detection's determinant/trace approximation, and discusses invariance properties (rotation, intensity shift) and limitations (scale sensitivity).
Sort: