Corner Detection

What Is Corner Detection?

Corner detection is a computational technique in image processing and computer vision that identifies points in an image where the intensity gradient changes significantly in two or more directions. Such points, called corners or interest points, correspond to physical features such as the junction of two edges, the tip of a textured region, or a sharp change in surface orientation. Because corners carry more local information than edge points or homogeneous regions, they serve as stable landmarks for matching features across images taken from different viewpoints, at different times, or under varying illumination.

The field draws from image gradient analysis, signal processing, and geometry. Corner detectors are evaluated by their repeatability under image transformations (translation, rotation, scale change, and affine deformation), their localization accuracy, and their computational cost, all of which determine suitability for a given application from offline 3D reconstruction to real-time robotics.

Gradient-Based Detection

The Harris corner detector, introduced by Chris Harris and Mike Stephens in 1988, remains the canonical gradient-based method. It constructs the structure tensor, a 2x2 matrix of averaged squared image gradient products in a local window, and classifies a pixel as a corner when both eigenvalues of this matrix are large, indicating strong gradient energy in two orthogonal directions. The Harris response function approximates the eigenvalue analysis through the determinant and trace of the matrix, avoiding explicit eigenvalue computation. OpenCV's Harris corner detection tutorial provides a widely used implementation reference and explains how the sensitivity parameter controls the trade-off between corner and edge response. The Shi-Tomasi detector (1994) modified the Harris criterion by selecting corners with the smaller eigenvalue exceeding a threshold, improving tracking performance in optical flow pipelines.

Segment-Test and Machine-Learned Detectors

The FAST (Features from Accelerated Segment Test) detector, developed by Edward Rosten and Tom Drummond in 2006, classifies a pixel as a corner by testing whether a contiguous arc of pixels on a surrounding circle of radius 3 (16 pixels total) are all brighter or all darker than the center pixel by at least a threshold. This arc test is far less computationally expensive than matrix-based methods, enabling real-time detection on embedded processors without dedicated hardware. FAST features are used in the ORB (Oriented FAST and Rotated BRIEF) descriptor, a royalty-free alternative to SIFT and SURF that combines FAST detection with a rotation-invariant binary descriptor. IEEE research on event-based Harris corner detection has extended these ideas to neuromorphic event cameras, which output asynchronous per-pixel brightness changes rather than full frames, enabling corner tracking at microsecond latency.

Relationship to Edge Detection and Motion Analysis

Corner detection and image edge detection share the gradient analysis foundation: edge detectors such as Canny locate pixels where gradient magnitude is locally maximal along the gradient direction, while corner detectors additionally require this gradient to be strong across multiple directions. Corners are therefore a subset of the broader class of edge junctions and terminations. In motion detection and optical flow estimation, corner points tracked across frames provide the displacement vectors used to estimate camera motion and scene structure. Algorithms such as the Kanade-Lucas-Tomasi (KLT) tracker select trackable features using a variant of the Harris criterion and estimate their motion between frames by iterative minimization of photometric error. The FAST and Harris fusion method in Mathematical Problems in Engineering illustrates how hybrid approaches combine the accuracy of Harris with the speed of FAST for detection tasks requiring both qualities.

Applications

Corner detection has applications across computer vision and robotics, including: