Am I Drunk?

Inspiration

The idea for this project arose during a general conversation we had yesterday. One of our teammate's family member had their life taken away by a person driving under the influence (DUI). But she's not alone -- her death joins nearly 10,000 people killed every year in the United States alone due to drinking and driving (that's about one person killed every 52 minutes). We devised a solution that gives you live, real-time feedback on your level of sobriety just by analyzing your facial expressions.

What it does

Am I Drunk? is a quick, less-than-10-seconds test to gauge if you're intoxicated. It uses a series of advanced, custom-coded algorithms to detect, isolate, and extract features from a short clip from your webcam. It then sends this off to two custom Deep Learning models for further inference. We use all of the outputs to finally display your level of intoxication. All you need is a decent camera (from a phone/computer) and access to the internet!

How we built it

Am I Drunk? uses React.js on the frontend, and Flask on the backend. The core component of the webapp is a series of Deep Learning models that we built in TensorFlow. We also utilized Google Cloud's App Engine to deploy the backend API. We chose this due to our need for speed and performance.

Challenges we ran into

Finding high-quality datasets for our models in the time we had was tricky. We sent emails requesting access to a couple of datasets, but settled on lower-quality, public ones due to limited time. This was also our first time working with Flask on the backend, so this was a fun learning experience for the both of us.

Accomplishments that we're proud of

We're proud of the amount we've been able to incorporate in the web app -- from training our Deep Learning models to learning about Flask. All this in the 12 hours we had to spend on this hackathon!

What we learned

The Flask and the GCloud API. Plus, Video Classification is much harder than it looks.

What's next for Am I Drunk?

Well-masked facial emotions can render our test with a false positive. This has led us to explore other features (such as vocal data) that could contribute to a more robust test. Currently, our prediction accuracy sits at around 85%, which is stellar in hackathon-amounts of time, but not enough to give the accurate results we need deployed in the field. We plan to scale this to a much larger platform so that anyone can use our (free) service to quickly gauge whether they're sober enough to sit behind a wheel, debate million-dollar deals, or operate heavy machinery. Had a similar project been used a few years ago by a certain beer-loving driver, our teammate's aunt would still be alive to continue being the gem she was in this world.

Built With

• flask
• google-cloud
• keras
• react
• tensorflow