Design of a Face Recognition System for Public Transport Fare Payment.

Abstract

Face Recognition technology is chiefly concerned with accurately re-identifying individuals through the use of mathematical face representations. Global advancements in the fields of artificial intelligence and computing have now made it possible for researchers to develop state of the art face recognition models with human level performance that can be leveraged to fix the gaps in a wide range of economic sectors.

This project introduces a fast, easy to use, and privacy oriented contactless payment system which to a great extent overcomes the deep lying flaws in the public transportation sector. The designed face recognition and internet of things based payment system puts an end to the poor passenger experience that has for long been associated with the public transportation sector in Uganda. Furthermore, the system also eliminates the current need for paper tickets during travel, which in turn accelerates the passenger boarding time and lowers the risk for the spread of communicable diseases such as COVID-19.

This project’s execution meticulously follows a well thought out set of system requirements specifications that are geared at developing an accurate deep learning based face recognition model, optimizing the model for real-time inference on the edge, developing a real-time payment system architecture and integrating the model within the architecture. Throughout the project’s execution, a total of 11,898 face images were collected locally from volunteers and also through web scrapping techniques to build a face recognition model. A Oneshot based face recognition model was the final model of choice after a rigorous model analysis and evaluation process. The Oneshot based model boasts of a recall of 97.35% and a 90.91% precision on unseen data and is a big improvement to the prevalent face recognition models due to its ability to thrive off single images. A real time payment system architecture, capable of completing transactions in less than 1 second, was also developed and its efficacy tested using the system’s test case specification.

Throughout the system’s design, key emphasis was placed on solving the prevalent problems in face recognition and deep learning systems i.e. privacy, security and latency. In regards to privacy, the system was designed not to store actual user images but rather to store encoded face strings, which would need to be uploaded only once for each user. The latency trap was beaten by ensuring that inference was carried out on the edge through hardware accelerators that ran the efficiently written programs. To solve the security problem, an anti-spoofing model, capable of distinguishing between real and fake faces was utilized. However, the development of an infallible anti-spoofing model was reserved for future attempts and iterations of this project.