Fast interest point extraction for augmented reality

摘要

The subject technology provides embodiments for performing fast corner detection in a given image for augmented reality applications. Embodiments disclose a high-speed test that examines intensities of pairs of pixels around a candidate center pixel. In one example, the examined pairs are comprised of pixels that are diametrically opposite ends of a circle formed with the candidate center pixel. Further, a pyramid of images including four rings of surrounding pixels is generated. An orientation of the pixels from the four rings are determined and a vector of discrete values of the pixels are provided. Next, a forest of trees are generated for the vector of discrete values corresponding to a descriptor for a first image. For a second image including a set of descriptors, approximate nearest neighbors are determined from the forest of tree representing closest matching descriptors from the first image.

主权项

1. A computer-implemented method, the method comprising:
identifying a candidate corner feature in a first image, the candidate corner feature including a set of pixels from the first image; determining a center pixel from the set of pixels; selecting, from the set of pixels, a subset of pixels, each pixel from the subset of pixels being at a respective distance from the center pixel, wherein the subset of pixels comprises a ring shape that surrounds the center pixel; downscaling the set of pixels from the first image, the downscaled set of pixels representing a second layer of a pyramid and the set of pixels representing a bottom layer of the pyramid; determining a first derivative value based at least in part on of a weighted sum of intensities of respective pixels from a right side of the downscaled set of pixels and intensities of respective pixels from the left side of the downscaled set of pixels; determining a second derivative value based at least in part on of a second weighted sum of intensities of respective pixels from an upper side of the downscaled set of pixels and intensities of respective pixels from a bottom side of the downscaled set of pixels; determining an orientation angle using at least an arctangent function on the first and second derivative values; shifting values of a subset of pixels from the downscaled set of pixels based at least in part on the orientation angle, the subset of pixels comprising a ring that surrounds a respective center pixel of the downscaled set of pixels; and confirming a corner feature of the first image based at least in part on:
determining that intensities of at least a pair of pixels are not within a threshold value of the intensity of the center pixel; anddetermining that an intensity of a first pixel from the at least the pair of pixels is unequal to an intensity of a second pixel from the at least the pair of pixels; and providing the confirmed corner feature to at least one object recognition process.