Self-driving cars use images from on-board cameras to navigate through cities. Research at West Virginia University could help solve a problem for those autonomous vehicles—recognizing the same image in different pictures.
Victor Fragoso, an assistant professor in the Lane Department of Computer Science and Electrical Engineering at WVU, said a computers inability to identify the same object in two different pictures is a fundamental problem in the field of computer vision.
“Humans have the ability to recognize the same object in different images, which allows them to understand and navigate through the three-dimensional structure of the world. This same task, however, does not come so easily to computers,” Fragoso said. “Computers need to understand a scene through an image, which is represented as a squared grid of numbers.”
Fragoso’s research, which is being funded by a two-year grant from the National Science Foundation, will measure the confidence the computer has when it determines that two objects with different viewpoints are present in one same scene.
“A self-driving car uses images from its cameras, among other data from various sensors, to localize where it is in a city. These cars recognize objects in the street and associate them with their respective 3D models,” Fragoso explained. “When this recognition process goes wrong, the car will estimate a wrong location and could even cause an accident.”
Fragoso will investigate ways in which the computer can reason about the different objects it detects and its confidence in accurately recognizing those objects. The proposed confidence measures can alert about wrong identifications, and can help the car trigger other mechanisms to avoid localization errors on-the-fly.
“Many systems perform slowly when the computer produces wrong object identifications in a pair of images,” Fragoso said. “The applications have to identify the incorrect object identification somehow, remove them and then use the correct identifications to operate.
“These confidence measures will not only help applications to prevent errors; they will also allow systems to speed up computations, and reduce the storage needed for visual content, such as 3D models of entire cities around the world,” Fragoso said.
Fragoso will investigate different ways to speed up computations in several applications in image stitching, panorama creation, augmented reality and self-driving cars, among others. Additionally, the project will investigate ways to compress 3D models while still ensuring a good functionality in many vision-based, navigation autonomous systems.
CONTACT: Mary C. Dillon, Statler College of Engineering
and Mineral Resources