Inspiration
Create something useful for the visually impaired and help them during a crisis. This hackathon of detecting fire using drones gave us an additional idea to build an app to help the blind navigate through fire.
What it does
When the code executes, the drone inside the simulation moves to the leftmost corner and starts moving autonomously throughout the scenario until it reaches the top/bottom right corner. Using the Faster RCNN model, trained using the Tensorflow API, the model is able to detect fire scenes from those images whenever the drone moves.
How I built it
Used the standard Tensorflow Object Detection API to train the model to detect fire in a particular image on the Microsoft Azure Virtual Machine. A Virtual Machine in Ubuntu 18.04 was set up for training the Deep Learning model.
Challenges I ran into
When we received a new set of images from the firedrone API, one of them was of an evening sky which at the first glance looked like the flames of fire. It was very easy for the Machine Learning model to get confused if not taken care of. Hence, we prepared a dataset which consisted of similar images viz. sunset and yellow bulbs. We didn't just want to detect yellow fire blobs because that had the chance of false detection. The aim was to detect the curves and features that the fire flames would have.
Accomplishments that I'm proud of
The dataset containing negative images of the fire helped to avoid false positive detection which would help us gain more points compared to others.
What I learned
It was a very fun and interesting learning experience especially dealing with the firedrone API and getting the images from the drone's FOV. Moreover, it was really exciting to automate the navigation of the drone and save all the FOV images which are processed to be used for detecting fire. Also, using negative images for training object detector was also an innovative approach to make our stand robust.
What's next for Firesome
This trained dataset along with Tensorflow Lite would be used to create a phone-based app, that would detect fires using the camera and give verbal guidance to avoid fire and navigate. This app would be extremely useful for the visually impaired who would be lost and stuck unless otherwise helped by someone under such critical situations.
Log in or sign up for Devpost to join the conversation.