Camera-guided sorting of recycled bottles on a conveyor

Mzansi PolyCycle buys baled bottles from collectors, sorts them, washes them and sells flake. Clear PET flake sells for about R9,500 a tonne at 97% purity or better, but only about R6,800 below that. Today 14 pickers per shift sort by hand from a conveyor, and the clear-PET stream averages 91% purity. The plant manager wants to test whether a camera and air jets could do the first sort.

The proposed station: - a conveyor belt 0.6 m wide running at 0.6 m/s, carrying up to 3 bottles a second in a single layer; - a camera over a lit section of the belt, capturing at 30 frames a second; - three air-jet banks 1.2 m, 1.5 m and 1.8 m downstream of the camera, blowing coloured PET, HDPE and contaminants into separate chutes, while clear PET stays on the belt.

The plant has no budget for hardware until the idea is proven.

You will: - set a timing budget from camera to air jet; - build a dataset from a public waste-bottle image set plus your own photos of bottles from home; - train a small image classifier and measure it with a confusion matrix; - simulate the conveyor, tracking and ejection in Python (or CoppeliaSim) to get purity and yield of each output stream; - tune the confidence threshold to trade purity against yield; - write a business case in rand.

You need no conveyor. Training runs on a laptop CPU or a free cloud notebook. A partner plant in Pune is also interested, so give the business case in rupees as well.

Done looks like a classifier with a confusion matrix on a held-out test set, a simulation that reports clear-PET purity and yield at 3 bottles a second, a chosen threshold, and a two-page report the plant manager can take to the owners. The target is 97% clear-PET purity with at least 85% of clear PET kept.