Michigan State University researchers received a four-year, $12 million federal grant from the U.S. Office of National Intelligence’s Advanced Research Projects Activity (IARPA) under its Biometric Recognition and Identification at Altitude and Range, or BRIAR, program.
The IARPA BRIAR program is a 48-month effort to deliver end-to-end software systems capable of detecting individuals at severe imaging conditions, extracting biometric signatures from the whole-body (such as an individual’s gait and/or body shape) and face, and fusing biometric information for robust multi-modal matching.
The MSU project entitled “Physics-driven Modeling and Learning for Person Recognition at a Distance and Altitude” is led by Xiaoming Liu, a research foundation professor at MSU; Arun Ross, the Martin J. Vanderploeg Endowed Professor; and Anil Jain, a university distinguished professor — all in the MSU College of Engineering.
Biometric recognition refers to technology for real-time and automated recognition of individuals based on their body attributes such as face, fingerprints, and iris. People use biometrics when they unlock their mobile phones using face or fingerprints, board international flights, and go through immigration after returning from an overseas trip.
In these applications, the camera or sensor needs to be near to the individual being recognized. But what about situations when law enforcement agencies want to find a missing person in the woods or locate a group of fugitives?
“When we combined the benefits of each of the three biometric modalities (face, body shape, and gait), we were able to achieve state-of-the-art accuracy as reported in the independent evaluation conducted by IARPA on sequestered data,” says Liu.
MSU’s system, called FarSight, accepts videos from drones as input and creates a candidate list of identities from a database. To build this prototype system, the MSU team had to address several challenges, including low-quality imagery, atmospheric turbulence, large variations in viewing angle and distance, and limited amount of operational data to train the recognition system.
“We have developed a successful prototype system that can look at one person from about 300 to 400 meters away,” says Liu. “During the next phase of the project, we’ll be looking at multiple people together at a range of 600 to 700 meters.”
As the team improves the range and accuracy of the system, it is also learning the relative strengths of various biometric measurements in images captured at different viewpoints.
“We were surprised to learn that a person’s height, build and gait often contain more useful biometric information than their face when capturing images from a long distance,” says Liu.
MSU’s team, which serves as the lead institution for this project, also includes Christopher Perry in the department of Computer Science, and Erin Bunting and Robert Goodwin in MSU’s Remote Sensing and Geographic Information Systems Center.
