Keypoints

The detection (extraction) of keypoints (keypoint detection), their characterization (feature description), and finally their comparison (matching) are closely related topics in computer vision. Applications using keypoints range from panorama creation to three-dimensional reconstruction, from visual odometry to object tracking, and in many other use cases.

The concept of a keypoint reflects the fact that not all image points, but only some of them, have a high probability of being identified unambiguously during matching. These points are distinctive, stable, and easy to detect. Over the last decade, as in almost every area of computer vision, major advances have been made in the development of local invariant features, namely keypoints that allow applications to define the local geometry of an image and encode it so that it is invariant to image transformations such as translation, rotation, scaling, and affine deformations.

This chapter discusses keypoint detection techniques. Point description and matching strategies are discussed in the following chapters.

A non-exhaustive list of algorithms for detecting keypoints is

Harris Corner

Harris formalizes the concepts of edges and corners from a mathematical standpoint and, by studying the eigenvalues of the covariance matrix in the neighborhood of a point, determines whether or not a corner is present. It is invariant to changes in illumination and to geometric transformations such as translations and rotations, and is minimally affected by scale changes (Section 6.2);

Shi-Tomasi

a variant of Harris that directly uses the minimum eigenvalue of the autocorrelation matrix to identify points that can be easily tracked over time; it forms the basis of the Kanade-Lucas-Tomasi tracker (Section 8.2);

AST

The class of Accelerated Segment Test algorithms (Section 6.5) identifies a keypoint by observing the intensity differences between points on a circle;

SIFT

analyzes the image at multiple resolutions and is invariant to similarity transformations (Section 6.3);

SURF

a more efficient variant of SIFT based on the integral image (Section 6.4).