This project develops a Micro-GPS system that provides centimeter-level accuracy and is reliable both indoors and out, based on specific landmarks in the "random" textures present in the world. The key idea is that all floors, such as the carpet in a building, the grain of a wood floor, the concrete on a sidewalk, and the asphalt on a road, have small imperfections, bumps, or variations in color from location to location. A downward-pointing camera mounted underneath a vehicle can observe specific, unique arrangements of these seemingly random variations, looking them up in an index to find out their precise position in the world. The developed technology can provide capabilities for better in-car navigations, such as accurate parking in a particular spot, pothole avoidance, and lane departure warning. Other applications might include smart wheelchairs that can stay on a sidewalk and avoid rough patches, scooters for the elderly and disabled, assistive technologies for the visually impaired, marker-free smart highways, smart robots in warehouses that can precisely position themselves next to shelves, and even domestic assistants that can handle day-to-day chores inside a home.
This research is based on a key idea that localization is possible based on specific features in the "random" textures present in the world: seemingly-heterogeneous textures that have unique variations everywhere but globally consistent image statistics. The key challenges of this project include developing methods for (1) detecting uncommon locations or "features" in a close-up image of the ground surface; (2) computing a feature descriptor for each detected landmark, in a way that is invariant to changes in orientation and lighting; (3) matching the features against a map: a pre-built database of features, their arrangements, and their locations in the world; and (4) being able to create and update the database to increase coverage and to account for changes. All of these are common components in contemporary systems for tracking, image alignment, and recognition. However, the individual algorithms have been tuned to work best for "natural" images. Instead, the project focuses on developing detectors, descriptors, matching algorithms, and update strategies that are tuned to the statistics of common ground textures. The research team investigates whether accuracy can be improved by combining descriptors based on color with ones based on surface normals or height fields; and the systems issues involved in scaling the system to widespread coverage.
