Improved facial recognition technologies inside cars

Institute’s researchers in biometrics developed tools for more efficient and more reliable facial recognition techniques in the automotive industry. They made available their results open source.

Facial recognition is a promising technology for the automotive industry. Applications range from controlling who is allowed to drive, for example adults in a family, to an instant adaptation of driving parameters to a driver’s specific needs. Technical challenges to integrate this technology to vehicles are very specific: limited lightning sources, relatively restricted on board computing resources, need for instant results and, of course, a high level of reliability. To ensure a reliable and robust face recognition, scientists from Idiap’s Biometrics security & privacy research group made two significant contributions.  On one side, they developed a “light” computational tool based on the so-called neural networks. On the other side, they created a biometrics database specifically for cars to improve sensors’ reliability.

Infrared sensors and public database

In a vehicle interior where lightning conditions can vary a lot, the solution to have good pictures of people faces is to have near-infrared sensors. To analyze in a reliable way these images, scientists usually use what we call artificial neural networks. This approach often needs many computational resources. « Our tests not only show that our algorithms are reliable, but that they also are fast and efficient in terms of computational needs to operate in real-time on a handheld device, such as a smart-phone,” Ketan Kotwal, researcher in the Biometrics security & privacy group, explains.

To ensure that their tool is reliable, scientists created a database of genuine and fake identification attempts in real-life conditions, that is, inside a car both in indoor and outdoor environments. Publicly available, data represents over 5,800 videos of 40 people filmed in various conditions and nearly 1,800 fraudulent identification attempts using, for example, paper and silicon masks or pictures and videos on a screen. “In addition to providing a tool that validates the genuineness of a face, we developed in parallel a database to test this tool more in depth and to establish a new standard in this field,” Sébastien Marcel, group leader, explains.

More applications in the near future

Being able to identify the driver of a car offers obvious advantages in terms of security and personalization of the driving experience. These on board technologies also have a potential for other applications. For example, they could ease access management for vehicle fleets or could be used to confirm the identity of the recipient of a delivery made by an autonomous vehicle. Such scenarios need reliable and affordable solutions similar to the developments made by researchers at Idiap.

More information

- Biometrics security & privacy research group

- Scientific publication “Domain-Specific Adaptation of CNN for Detecting Face Presentation Attacks in NIR” 

- Public database