Millions of people move through the Nation’s airports, mass transit system, stadiums and other public spaces every day. Every tool available for keeping travelers and the public safe is necessary to help paint a full picture of the security situation. When it comes to vetting large crowds, face recognition technology is maturing quickly and could make a big difference for law enforcement and the Homeland Security Enterprise.
The use of face detection and recognition has the potential to contribute to the capabilities of future law enforcement applications, such as high volume screening in crowded places, low volume forensic examinations, and crime scene investigations. The full extent of face recognition technology’s accuracy has not been independently tested—until now.
The Department of Homeland Security (DHS) Science and Technology Directorate’s (S&T), Face in Video Evaluation (FIVE) Program provides an independent evaluation of face recognition algorithms for various law enforcement applications. The findings are highlighted in the FIVE Program report which was released in March 2017.
“The FIVE Program focuses on assessing the capability of face recognition algorithms to correctly identify people appearing in video sequences. Both comparative and absolute accuracy measures are vital to determining which algorithms are most effective for a given application and whether any are viable for conducting security screening operations,” said S&T First Responders Group Office for Public Safety Research Program Manager, Patricia Wolfhope.
Current face detection and recognition algorithms work best on cooperative subjects. Non-cooperative face recognition means that the person does not need to cooperate with the cameras and therefore, their face could be off angle, off perspective, various resolutions, poor lighting and the face could be obscured. Because of these factors, non-cooperative face recognition algorithms tend to have decreased performance than if the person was cooperating with the cameras. Non-cooperative face recognition systems are meant to augment current operators and aide in speeding up throughput effectively.
Obtaining video data that is indicative of law enforcement applications was key for the FIVE Program to perform independent evaluation testing. FRG funded the initial data collection in 2013 at the Toyota Center in Kennewick, Washington. The objective of the data collection was to collect video data that simulated realistic crowd flows in public spaces along with both straight and winding queuing lines. The testing used volunteers.
The venue was a multipurpose stadium with a seating capacity of 6000 people. Three different locations inside the Toyota Center were chosen to represent the selected operational settings. The three operational settings included 1) unidirectional crowd flow (captured by cameras located at the main entrance), 2) bi-directional crowd flow (captured by cameras in the hallway), and 3) winding queue (captured by cameras at a beverage stand).
For unidirectional flow, three video camera locations were selected above the main entrance doors in the foyer. Four cameras were set up at the queue and four in the hallway to capture bidirectional flow. These types of tests occur only in controlled environments with volunteers to validate the technology being evaluated.
Additional video and camera positioning data were contributed by HSARPA’s Apex Air Entry/Exit Re-engineering (AEER) program – now BMD Port of Entry People Screening. AEER’s Maryland Test Facility (MdTF) was instrumented to collect real-time performance and human factors data while volunteers work their way through simulated operational conditions commonly encountered in airports and other high-throughput environments like checkpoints and ports of entry.
S&T worked with the National Institute of Standards and Technology to develop an independent evaluation of 36 face recognition algorithms against the dataset using various watch list configurations and camera outputs. The FIVE report is provided to aide law enforcement in purchasing the right algorithm for their specific applications and placing cameras in the best position to obtain the highest face detection and recognition performance.
“We are constantly making significant advances in recognition technology, thanks to our cooperation with partner agencies and the private sector”, noted Wolfhope.
For more information about S&T’s work at the Maryland Test Facility, please contact firstname.lastname@example.org.