FireMe

Landing page
Info about our project on the landing page
Testing the camera feed with a fire
Receiving alerts about a fire with information on the location
Signature meme for our project

Inspiration

Climate change or poor policy? As Australia's wildfires see some relief, the blame game ascends. Well, it's 2020, and the only thing we should blame is the lack of technology to prevent disasters like this. The plight of dozens of animals being treated for burned paws and singed fur is raising fears.

The questions raised is, why weren't animals relocated or fire extinguished when it was just growing. Baffled by these questions, which kept disturbing our minds for the past few weeks, we utilized the opportunity, and the platform Hack Arizona provided to develop something to help prevent something similar happening in the future.

What it does

We started with the basic thought that the world is filled with cameras or CCTV. China has created a social credit system that utilizes and recognizes people's faces to evaluate their social score, and reward/penalize them based on that.

Our idea is based on the fact that we could use those very same cameras, track the videos from them, and alert users if a fire is detected. Once a potential fire is detected, the designated/registered users are sent a notification. This fundamental idea can be scaled to an almost infinite number of cameras or even mobile phones, which we use as CCTV cameras for demo purposes. It can also be used for indoor houses or store monitoring, where fires are generally caused by passive short circuits. After all, you wouldn't want to return home to see it burnt.

How I built it

We first use feed from the cameras and process it by splitting it into the frames. We use the camera from a smartphone for demo purposes, where we access its feed through a secured ipv6 connection by using its to stream the video. After obtaining the feed, we test the frames using a deep learning-based model trained on a massive fire-based image dataset. The frames are tested as API calls to the backend server, which runs the model. The model is generated using TensorFlow v2 with more than 100 layers of convolutional filters. The model was able to detect fires with an accuracy of over 91.1%.

Once the fire is detected, we send SMS alerts to the registered users using the Twilio API. Since we use web service based API calls, our model is extremely scalable, both with several cameras as well as several users.

Challenges I ran into

We had issues choosing the right stack and thus had to go back and forth with javascript and python as well as code for updated versions in some instances where old python modules were deprecated.

Accomplishments that I'm proud of

One colossal accomplishment would be the development of a completely scalable prototype by using deep learning for predictions and rts for IP based streaming. We are proud that we were able to do this in a limited time frame. We are also happy to share/boast that our application would actually be useful and may even save lives.

What I learned

We have learned a lot more than what we learn in classrooms in a typical Fall semester. We learned the cross-use of python and javascript front-end and backend servers as well as web service-based development. We learned how to optimize programs to run realtime. We also learned to utilize both limited time and resources we had.

What's next for Fire Me

Our next step would be to integrate the software to real-world CCTV cameras. We would also like to gather a better dataset for training the deep learning model as well as make sure it is accurate. We would also further like to optimize the project to make it run on even low-end servers/computers.