Human identity and understanding of person behavior using smartphone devices

Purpose and goal

A fully working CNN for face recognition has been developed, with the restriction of operating under limited hardware. The system is also designed resilient to face rotation, a frequent perturbation in mobile environments. A paper to a prestigious conference (ICPR) co-written with the company has been submitted, describing the system. The model obtained is of just 4.4Mb, and is trained on state-of-the-art databases (>6 million images). As a reference, regular face recognition models have sizes exceeding 100 Mb, which are unsuitable for embedding in Android or IOS applications.

Expected results and effects

Findings will enable increased security and enhanced user experience in the exploding smartphone market. The project has also allowed new cooperation between the Swedish academia and a European SME that provides smartphone-based biometrics AI solutions (FacePhi), in the context of a wider effort to establish a consortium for an international funding application (to EU H2020, or other). This project has also increased the understanding by the applicant of issues, challenges and constraints associated with market implementations of biometric solutions.

Planned approach and implementation

The proposal involves a company (FacePhi) who has brought know-how on implementation of AI solutions in smartphones, and an academic partner, who has brought knowledge of latest research in identity and behaviour understanding. By hands-on experience, the applicant of the Swedish academia has assessed how an AI research solution is transferred to a product, taking into account inputs and needs of end-users, or hardware restrictions. Mobility was carried out in 2 periods of 1.5 months, aiming for extended cooperation that goes beyond the duration of the mobility