Safe.AI | Devpost

Safe.AI logo

Inspiration

With the rise of deadly mass shootings in schools, something must change fast. Regardless of political views, there is one thing we can all agree on: we need more security in our schools. Since the Parkland Massacre in 2018, there has been, on average, a school shooting every 12 days. Something must change.

What it does

Using object recognition, Safe.AI is able to detect a gun in a provided video security feed. It is then able to alert authorities with the time, date and location of the detection for immediate response. Current schools have poor surveillance systems, most having huge delays in playback.

How we built it

Using the OpenCV library in C++, we first began by collecting data sets to create a classifier in order to detect a gun. We needed as many positive images (i.e. pictures of guns) as possible to increase our accuracy. We also needed just as many negative images, which are images that do not contain the object, for the library to learn how to recognize a gun, similar to how humans learn.

Challenges we ran into

We quickly realized that training large amounts of data, especially at a high resolution, requires a lot of time and processing power. Due to this, we had to quickly find ways to create classifiers efficiently so we could have time to test and make any adjustments if needed.

Accomplishments that we're proud of

We are very proud that in less than 24 hours, we were able to train a classifier and successfully detect a gun with no prior experience in image recognition. We are also proud that for every problem that occurred, we worked together as a team to solve it so we could move forward in the development project.

What we learned

We learned that just because a task may seem simple, there is always things that could go wrong, and that we must expect the unexpected. We also learnt it is very important to research library documentation beforehand to learn our limitations using the library.

What's next for Safe.AI

Our goal in the near future is to add as many data samples to our set as possible to increase accuracy and to avoid unwanted situations with the recognition. We also want to implement more recent and updated image detection libraries and APIs to help Safe.AI be the best tool for our society.

Built With

Submitted to

HackDartmouth V

Created by

I worked on the .xml files, selecting the data fed into the deep learning cascade generator for the AI to have the necessary information to train (positive and negative images). A lot of trial and error analysis was needed for this optimisation part. It was my first time using the Open-CV library and in the futur I believe it would be more practical to use Tensor Flow or IBM Watson

Thomas Tran
Student in Computer Engineering at Concordia University
I worked implementing the HAAR Cascade to detect and surround the weapon with a rectangle for identification. I also worked on providing notifications and alerts of the recorded detection while creating a "confidence" generation algorithm.

Michael Nareg Arabian
Computer Engineering @Concordia Univeristy