Keypoint Detection by Cascaded FAST
Matching of keypoints between images is used in applications such as specific object recognition, panorama image generation, and image searching, and is an important process in the field of computer vision. Since keypoint detection in such applications is a task that is equivalent to preprocessing, it is considered necessary to have processing that is as fast as possible.
A method called Features from Accelerated Segment Test (FAST) has been proposed as a method of rapidly detecting corner points in an image. FAST constructs a decision tree through machine learning and detects corners rapidly by traversing that decision tree. However, there is a problem with detecting large numbers of corner points from natural regions with complicated textures (such as tree leaves or shrubbery). Such corner points are unwanted in the matching of keypoints and increase the cost of keypoint matching calculations.
We therefore propose Cascaded FAST in which three types of decision tree that reference different boundary lengths are aligned in a cascade. Cascaded FAST makes it possible to rapidly detect points that seem to be corners, while inhibiting the detection of unwanted corner points.
Cascaded FAST learns three decision trees with different reference pixels, in a similar manner to FAST, and rapidly detects corner-like points by arranging these decision trees in a cascade. During this time, it evaluates corner-likeness by computing orientations. It can also obtain the scale and orientation of corner points by using pyramid processing to obtain scale. This has made it possible to implement rapid keypoint matching by describing binary features by ORB in patch images that have been normalized by using the detected scale and orientation. In comparison with the processing time for keypoint matching of previous methods, we have greatly reduced the distance calculations when searching for corresponding points, enabling matching at a speed of approximately 20 fps.