AST

The final class of keypoint detectors is known as the Accelerated Segment Test, developed by Rosten. Three slightly different versions of this algorithm are currently available.

Figure 6.5: FAST: the 16 pixels on the circle of radius 3 used to test for consecutive pixels.

The first version of Features from Accelerated Segment Test FAST (RD05) is probably the most intuitive: points are considered distinctive if they have a contiguous sequence of $n$ pixels along a circle of a given radius, all brighter (or darker) than the central pixel, which serves as the reference for gray level. For example, FAST-9 analyzes the 16 pixels on a circle of radius 3 and checks whether 9 consecutive pixels are all above or all below a certain threshold relative to the central pixel. In subsequent versions (RD06), extraction is optimized using decision trees trained to identify distinctive points that maximize the local information content. These trees always process the pixels on the circle.

This approach has become typical in recent years: with the abundance of public datasets, classifiers have been widely used to build detectors of stable keypoints. Given primitives that describe a point's neighborhood, an optimization technique can identify those that are most stable for a particular task. Rosten's paper also provides an excellent survey of earlier keypoint detection techniques.

The final variant, FAST-ER, extends the analysis region beyond the pixels on a circle to include all pixels in the neighborhood of the central point.