Search giant looks to use invention for video processing in mapping, machine vision, among other applications
Google Inc.’s patent for appearance augmented 3D point clouds for trajectory and camera localization, an invention to be used for video processing in mapping, machine vision, computer applications, and the Internet, was approved by the federal government this week.
Patent No. 8,442,307, filed two years ago, is based on the premise that accurate geographic position information is necessary for several applications – from map building to geo tagging data to localizing images. The patent was granted approval May 14 by the U.S. Patent and Trademark Office and listed the inventors as Roy Anati of Philadephia and Dragomir Anguelov and David Martin, both of San Francisco.
Here’s the patent’s rationale:
“An essential component of any [geographic positioning]system is the ability to correctly identify the position of an object of interest where that position is estimated either by multiple sources or by a single source at different times. By identifying the position of multiple objects in relationship to a moving platform, a trajectory path of the moving platform can be established.
“However, if trajectory paths are estimated either by multiple sources or by a single source at different times, the multiple trajectories must be aligned to be combined. Once multiple trajectories are combined into a single coherent whole, subsequent applications use of the trajectory data is greatly simplified.
“Approaches for matching an image to an image set to align multiple trajectories typically focus on extracting scale-invariant feature descriptors from the image collection and constructing a fast query index over the descriptors.
“The problem of matching an image to an image set typically includes extracting the features from the query image and performing a look-up in the index for the image with the most similar and geometrically consistent set of feature matches. However, the problem of finding a consistent pose for moving (camera) platforms at trajectory intersections makes this matching difficult due to loop-closing.
The invention, the patent read, addresses the problems of “finding a consistent pose for moving platforms at trajectory intersections and matching an image to an image set.”
And here’s how it does it. The appearance augmented 3D point cloud receives the first and second posed image, extracts image features from the first and second posed image, and compares the extracted image features from the first posed image with the extracted image features from the second posed image.
“Next, a comparison is done to identify one or more matched features of the first and second posed images based on a feature appearance, where a grouping of the matched features of the first and second posed images is performed where the grouped matched features are associated with a first 3D point.
“A position of the first 3D point is identified based on a positional triangulation of the grouped matched features associated with the first 3D point, and the first 3D point is then augmented with the grouped matched features.”
The newly approved patent borrows from a 2012 Google patent for an invention that uses image and laser constraints to obtain consistent and improved pose estimates in vehicle pose databases.