Inspiration

  • Everyone can contribute to helping the environment by recycling at almost no cost, but this is limited by lack of knowledge of recyclability of waste items
  • Many smart bin solutions based on object detection and item separation have high set up cost, or inconveniences users with sensor and mechanical lags
  • Aim is to better inform users about recyclability of waste items, without becoming obtrusive, disruptive or inconvenient
  • Build better recycling habits through encouragement and positive incentivisation
  • Companies are keen to support eco-friendly initiatives – can promote partnerships with companies through reward schemes
  • New waste items/packaging arrive daily, so traditional rulebased classification methods not sufficient. We need a generalised data-driven method (hence DNN).

What it does

  • Hardware setup: user mounted device, cheap, easy to install device on the inside of a bin lid or rim
  • User signs up to Recyclify service/download app
  • after each waste item is thrown in the bin, device keeps track of whether items are correctly recycled or not....
  • User received push notifications if items are incorrectly disposed with a gentle reminder for the next time, and encouraging messages for
  • user accumulates reward points based on consistent/consecutive recycling habits
  • reward points can be redeemed for discount vouchers/loyalty points with partnered companies

How It’s Built

  • Raspberry Pi based camera captures and uploads an image stream to a Microsoft Azure hosted container
  • Server monitors incoming images, detecting when new objects enter using a structural similarity index between frames
  • Flagged images are passed to a DNN to classify the type of product and recyclability
  • Updates SQL based customer reward points database hosted by Azure and apushes notifications or emails when a reward is earned or item(s) are incorrectly recycled/binned

Challenges Encountered

  • Hardware reliability was a major challenge, with the Raspberry Pi
  • Sensitivity of ML model to natural variations in image quality/lighting etc.

Accomplishments that I'm proud of

  • Able to produce a proof of concept end-to-end pipeline
  • Tackle shortcomings of ML models with classical image processing

What I learned

  • Azure container, SQL database and VM
  • Circumventing limitations in hardware and data availability

What's next for Recyclify

  • ios/web-app development
  • Approaching business partnerships (reward points, vouchers etc.)
  • Accumulate further data and more training to predict the recyclability of out-of-training samples
  • Refining the hardware design to for easy mounting, space and cost efficiency

Built With

Share this project:

Updates