Canterbury Bankstown

Local councils around Australia face the ongoing challenge of residents placing non recyclable items in recycling bins. Mixed waste adds significantly to the cost of waste processing.

Broad based education campaigns are effective to a point, but what ultimately shifts behaviour is the inspection of the contents of an individual resident’s bin. Unfortunately regular manual inspections prior to collection are both expensive to perform and limited in their effectiveness as you can only see what’s at the top of the bin.

The Challenge

Canterbury Bankstown, an innovative Sydney based council, saw the opportunity to move beyond traditional techniques for minimising contaminated waste and instead use Artificial Intelligence to identify contaminated waste at the moment it’s collected. Such information can then drive targeted consumer education campaigns. 

Canterbury Bankstown engaged Alliance Software to create a series of proof of concept Machine Learning models to confirm the effectiveness of AI in waste contamination identification.

The technical challenges to achieve this were significant including:

• Grouped hopper imagery

  Due to the configuration of garbage collection trucks, footage can only be taken in the ‘hopper’ (the large container that holds waste), not as it is falling through the air. This means our first challenge was to create models to identify each resident’s individual bin contents within the larger hopper.

• Waste comes in all shapes & sizes

  Both recyclable and non-recyclable waste comes in all shapes and sizes. Indeed, even human reviewers often find it difficult to determine what’s recyclable or not from video footage.

It should be noted that Canterbury Bankstown intend to use human reviewers prior to notifying residents of inappropriate waste. The goal of the system is simply to present these council staff with selected imagery from the many hours of footage that show, with high probability, the presence of non-recyclable waste.

Our Approach

We worked closely with Canterbury Bankstown to creatively tackle the challenges of the project. Techniques applied included:

• Lid identification

  In addition to creating Machine Learning models to categorise waste types, we also built a model to identify each individual bin collection from a video stream. It turns out modern ML models do an excellent job of detecting large yellow bin lids as they swing past the camera.

• Label consistency

  Due to the high variance of shapes, sizes and colours within waste, it’s common for two human reviewers to categorise the presence of contaminated waste differently. This makes training difficult and so to increase accuracy, we ultimately used three different human reviewers to review the same footage. This gave us higher quality training data and increased the confidence of models.

• Training by waste type

  Rather than trying to simply identify recyclable vs non recyclable waste as two broad categories, we actively worked to identify the presence of a range of specific waste types. We found we were able to classify some classes of wastes with greater accuracy than others. For example plastic bottles are easier to classify accurately than plastic wrap.

The Results

The project was a great success, even being featured on Channel 9. Based on an initial training set of 4,000 images, we created Machine Learning models able to predict the presence of common contaminants 80% of the time. As an input into a human review process, this rate of accuracy is already sufficient to proceed with broader deployment.

In addition, we’re confident that with more training data and further model refinement, both our accuracy and the range of contaminants we can detect will increase.