Inspiration

On May 25th, George Floyd was brutally killed by policemen. His death has enlightened thousands of people, marking the existence of fatal amounts of racism and the trend towards the United States becoming a police state.

As people bypass quarantine and curfew laws to protest against this injustice, several new cases of police brutality have come up. Newsfeeds of homeless people being hit by rubber bullets and a small girl being pepper-sprayed has continued to fuel the pathos of many people. When we see individual stories of the atrocities committed by our own protectors, there are usually two trains of thought. Many people are enraged and take to the streets more often. Others, although horrified, aren’t convinced to protest because they’re only exposed to a few news articles with invigorating headlines but a more dull truth.

However, my team’s research has shown that there are a lot more cases of brutality than we see. These are a few stills of videos we found that were buried deep in the internet. During every protest, several new videos surface with unprovoked action against protestors, but only a few go viral. We immediately realized the importance of visualizing this data. Emotion is strong, but data is stronger! With this information, we can convince more people to protest against injustice, this time armed with facts and logic.

What it does

With a complex architecture, our team set out on an ambitious goal to try and map out all cases of police brutality. Knowing that most of these videos surface on Reddit, we built a metadata Reddit scraper that searches for media files. Moreover, we also built an application with React Native (thus making it cross-platform!) that allows users of our app to post their own photos and videos that they've recorded onto our app.

We prioritized the safety of the protestors from a cybersecurity standpoint since we want all users to stay anonymous. Thus, we built a unique CNN model from TensorFlow that allows face blur from images and protects the identity of any people that might be on the video. We don't want people to get targetted!

However, we don't want to just append any random media file to our database! Using Amazon's S3 Bucket and Rekognition APIs, we were successfully able to extract how much violence was in our video. Then, we uploaded these files to DigitalOcean, extracted a file location, and stored this information in MongoDB along with the case's geolocation!

On the client-side, the user is able to see a feed of the most recent cases from our MongoDB, and a map showing the geolocation of every user-uploaded case.

What's Next for ActivistArmor?

In order to accomplish our goal of arming activists with information, we hope to add a web app to our project, get our app published on the App Store and Google Play, and expand our scraper's abilities in order to collect content from Twitter, Facebook, websites, and more.

A problem we face with this version of our app, is that our server cannot handle video well enough to blur faces in the videos, however the code and functionality to do so exists in our program already. If we raised enough money, as little as $100, we could upgrade our AWS server to include a GPU. This would protect protestors from having their face appear in videos on our website to protect them from being prosecuted by authorities.

Another way we would like to improve ActivistArmor is to find videos of the same event from different angles. We could do this by implementing object detection ML and finding match objects in videos. Protestors could then see well-documented video of police brutality from multiple angles.

Lastly, in order to speed up our content flow, we need to implement a way to remove duplicate videos and crop videos down to the content that matters. This will allow us to utilize any grant money or funds we receive in a more responsible manner and will work better for protestors by getting them only the content they want as quickly as possible.

Built With

• ai
• amazon-web-services
• cnn
• digitalocean
• galio-framework
• javascript
• ml
• mongodb
• react-native
• rekognition