Computer Vision Machine Learning Recycling Can

Recycling Can

Overview

Recycling contamination happens when non-recyclable objects get introduced into the recycling system. This is a major issue in the recycling process where a few incorrectly recycled objects can cause an entire batch of materials to be diverted to the landfill instead of being recycled. Current processes use a centralized system at the recycling facility to sort the recycled materials (often manually) after it has been mixed together, but this is very costly and prone to error. We propose a decentralized system where the sorting happens at the recycling bin to prevent mixing from happening in the first place. This is an affordable, automated, and reliable solution that can save tons of recyclable materials from going to landfills due to recycling contamination, leading to a more sustainable future.

What it does

Our device is a recycling bin that uses a camera and machine learning-powered computer vision to determine if an object can be recycled. If it can be recycled, the bin opens its lid to allow the user to throw in the object. We planned to also reward the user for recycling by potentially giving them a blockchain currency using Pi Network's platform, but we ran out of time to implement this.

How we built it

Our system uses a Raspberry Pi to take images from a camera plugged into it. The device then sends the image to the Google Cloud Vision API to classify objects in front of the camera. Once it determines if the object in front of the camera can be recycled, it will send a signal to a servo motor to open the recycling bin's lid to allow the user to throw the object into it.

Challenges

We originally tried to use OpenCV running YOLO (You Only Look Once) Object Detection with a model trained on the TACO (Trash Annotations in COntext) dataset to determine what type of trash is being shown to the camera. However, the Raspberry Pi had difficulties running it and we decided that the power consumption would be too high to do the processing on-board. So instead we decided to use the Google Cloud Vision API, made possible with the Google Cloud credits provided to us for the hackathon.