Inspiration

Ever witnessed individuals on campus throwing their humongous bag of trash into the "General Trash" bin, without throwing each type of items (plastic, glass, metal) into their respective compartments?

As a result of these individual actions, it is upto Serco to compartmentalize your trash, painstakingly going through each and every trash item, risking both their health and physical hygiene.

While we can raise awareness about this issue and prevent people to do this, we can only do it till a certain point. It is high time we leave it upto Artificial Intelligence to take over this matter, once and for all.

What it does

Our project, Trash Sorter Extraordinaire, is a software which efficiently detects the different types of waste in a pile of random trash through feeding it pictures or videos. Think of how Serco digs into a heap of trash to find the plastics, metals and glass and sort them accordingly. Now, Serco will simply take a picture of the trash, send it to the Trash Sorter Extraordinaire software, and it will indicate which items are plastic, metal and glass. With this knowledge, the only job Serco needs to do is put the trash items into their corresponding bins — plastic, metal, glass.

How we built it

In order to implement this software, we used jetson nano as an accelerated gpu device to incorporate transfer learning, and also makes it fast and efficient. We utilized nvidia's resources on github as a platform for us to use AI applications using jetson nano, as it provides similar tutorials and examples. Next, we had to use a pre-existing model and train it to our specific dataset, and as nvidia's training script is the gold standard to train datasets using transfer learning, we were able to use it for our own dataset as well. Then we created the object detection program utilizing pytorch and jetson-inference python libraries, and we used an open-source website to download sample images in order to test our program.

Challenges we ran into

Jetson-nano repeatedly ran into errors during training. The error were sourced from the build of the gpu, and hence we were able to fix it by clearing the cache and restarting the training process entirely.

Accomplishments that we're proud of

We're proud of the degree of accuracy to which the detection program works, which indicates how well the training program worked on our dataset. For example, the software was able to detect most glass bottle with an accuracy around 85-95%.

What we learned

Through nvidia's online resources, we were able to learn a great deal about using artificial intelligence applications in jetson nano.

What's next for Trash Sorter Extraordinaire

Currently, the Trash Sorter Extraordinaire is able to detect both glass and plastic, and we aim to extend it to all categories of waste, including metal, bio-degradable, etc. We prioritized detecting plastic objects as plastic is non-biodegradable and is the most harmful waste to the environment. In terms of future research, we plan to extend the Trash Sorter Extraordinaire in the scope of robotics. Using this software, the robot will be able to pick the trash and compartmentalize it according to the type of trash which this program detects. It will not only eliminate the need for Serco to collect and allocate trash, but it also has the potential to reduce litter on the NYUAD campus.

Built With

Share this project:

Updates