Mask Mandator

Project introduction

Mask Mandator is a campus-wide system that enforces the proper usage of PPE, like masks, before entry into common areas and other population dense buildings.

Issue/Inspiration

Mask Mandator was inspired by the threat of the COVID-19 pandemic back in November 2020. Despite the US breaking top COVID infection statistics in the past, many people did not wish to follow simple guidelines for everyone’s safety. Seeing those people without masks on campus made students and staff feel quite uncomfortable, and we wished that there were more preventative measures in place. Our desire to aid in the fight against the pandemic and protect people inspired us to launch this project.

What it does

We combine already widely used card and ID systems with cameras and AI to detect if someone is properly protecting themselves and others by wearing a mask before entering rooms. The system checks with Google Cloud’s BigQuery database to match IDs with data points such as if they're immunocompromised and the number of mask infractions they've had in the past and act accordingly to grant or reject the person access.

How did your project evolve with the support of the COVID-19 hackathon fund by Google Cloud?

The COVID-19 Hackathon Fund by Google Cloud was the reason this project continued after HackRPI. The fund allowed us to get creative and explore the powerful tools that Google Cloud offered without fear of accidentally charging our credit cards. As a result, we were able to experiment with a lot more technologies than we were originally able to! We also got heaps of support from our mentor, Jerry Ye .

How you built it

After an individual scans their identification card, and the camera scans for a face. Transformations occur to pre-process the image. One such transformation is using facial detection from Google’s VisionAPI to crop the image to the user's face. This preprocessing ensures we get accurate results when we try and predict whether or not the person is wearing PPE. Then, the image is tested with the Google AutoML model, which returns whether the individual is wearing a mask or not. If the individual is wearing a mask, the door unlocks and allows the user through. If the model does not detect a mask, the door stays locked and an infraction is recorded.

Challenges you ran into

Our first implementation of Mask Mandator at HackRPI was far from perfect. To make it work in the limited time frame, we had to cut a lot of corners. This included making the program based around a Python GUI, and we had the original identification method based around holding QR codes up to the camera. This was far from how we imagined Mask Mandator would work in the real world, and we had to come up with new ideas in this regard to fit our vision better. We also ran into some issues with our AutoML model, likely due to some bias in the data. It didn’t run the best under certain conditions, so we had to use a variety of image preprocessing techniques to get a higher accuracy. Lastly, our database from HackRPI didn’t fit in well with our Google Cloud tech stack. So we had to swap it out for BigQuery.

Accomplishments you are proud of

We’re proud of how far we were able to come as a team. We are all first-term freshmen at Purdue University, and it has been exciting working with one group for a long term project like this. We’re all proud that we were able to get our physical model up and running, and getting it to a polished state similar to what we envisioned when we first started HackRPI. We were also able to create a polished product website to showcase our project! https://mask-mandator.github.io/

What you learned

We all learned a lot about having to manage our time and resources across multiple months to build Mask Mandator. This is the first long-term extra curricular project for all of us. On the more technical side, we learned a lot about how preprocessing images can improve the performance of models. Paying attention to data biases are important when you’re initially training, but even with a somewhat biased dataset you can sometimes get around it with clever transformations.

What’s next for your project

The next steps for Mask Mandator would be hooking up the physical model to a real ID and door locking system. As well as making a more cost effective and polished model.