We created a field sobriety test web-app harnessing the power of AI to increase accuracy & accessibility, and to reduce bias & cost of a normal field sobriety test. Built with a WixCode frontend and two artificial neural networks (one custom-made, custom-trained convolutional neural net model) on the backend, our web-app is a quick & easy exam to determine whether or not someone is inebriated. Driving under the influence is an important societal problem that needs tackling. Whether it is used by friends trying to convince others not to drive inebriated, by law enforcement to reduce bias and increase efficiency or by employers seeking to keep their employees sober on the clock, sober.AI has the potential to have a huge impact while saving lives.
With the legalization of cannabis recently in Canada and the ever-lasting threat of people endangering lives while driving under the influence, we wanted to make an accurate, easy-to-use, yet powerful tool to detect sobriety in individuals. This could be used as a tool to prevent your friends from driving under the influence, by law enforcement to detect inebriated people or companies looking to make sure their employees are sober before operating heavy machinerie.
What it does
A short 3-step exam enables you to accurately assess if the person you are doing the test on is indeed sober, or if this person is inebriated. A percentage of confidence of sobriety will be given. These tests are the ones most widely used by law enforcement in North America for an accurate assessment.
How we built it
We built this web-app starting with a WixCode front-end, that communicates with an Azure VM that will detect sobriety through computer vision. One of the tests uses a custom-built, custom-trained TensorFlow CNN model. The other uses Azure's custom vision classifier. We leveraged NC6's immense server power to effectively train our neural networks very quickly, with our custom-created data-sets, to achieve stellar results for an accurate result, all in hackathon-amounts of time.
Challenges we ran into
Getting the proper versions of Cudnn to play nicely with TensorFlow. Getting data-sets of images of people who were high.
Accomplishments that we're proud of
Training a custom-built artificial neural network with custom-made data-sets, all in less than 36 hours! WixCode integration for our front-end.
What we learned
WixCode syntax and API calls, Azure services, classifying videos is harder than it looks.
What's next for sober.AI
We would love to increase the accuracy even more, and integration into law enforcement field sobriety tests for more accurate results. Also, we could make it into a widely used app that people use to check on their friends before they drive. Also, it can be used in hardware or big-box stores to test people's sobriety before they use heavy machinery, which when operated under the influence is very dangerous for customers/employees and illegal. We also want to implement different instances through sessions or tokens to make the web-app more scaleable.