Smart-Bin Waste Classification
IoT-enabled bin with a CNN classifier (Xception-inspired) for real-time sorting of plastic, metal, and organic waste, plus a cloud-analytics layer for collection-route optimization.
/ Outcomes
- 87% accuracy on three-way classification (plastic / metal / organic)
- End-to-end IoT → ML → cloud pipeline running on the bin itself
- Suggested workflow change with potential to improve collection efficiency by ~25%
Overview
A smart bin that decides what kind of waste it just received and surfaces that signal up to a cloud-analytics layer that can re-plan collection routes. The classification model is a CNN inspired by Xception, fine-tuned for a three-class problem (plastic, metal, organic). The deployment target was a constrained IoT board, not a server.
Approach
- Xception-inspired backbone for separable-convolution efficiency. The depthwise-separable structure of Xception was the right shape for an IoT inference target — it kept the parameter count and per-frame compute low enough to run on the bin.
- Three-class problem, narrow but representative dataset. Plastic, metal, and organic cover the bulk of household waste streams. Training data was scoped tightly to keep the classifier accurate on the actual deployment distribution.
- End-to-end pipeline, not just a classifier. Sensor capture → on-device inference → cloud analytics. The classifier output is only useful if the data flows somewhere that can act on it.
What I built
- The CNN classifier (Xception-inspired) in Python + TensorFlow, trained against the three-class waste dataset
- The IoT integration: capture, on-device inference, and the publish path into the cloud layer
- The cloud-analytics layer that aggregated bin events into a candidate collection-route workflow
- The proposed workflow change derived from the route data — measured in a simulation against the existing collection schedule
Results
- 87% accuracy on three-way real-time classification.
- A working end-to-end pipeline from sensor capture through cloud analytics — proof that the constrained-device inference path was viable.
- The simulated workflow change suggested ~25% improvement in collection efficiency versus the existing fixed-schedule approach.
Lessons
The classifier was the smallest part of this build. Most of the engineering went into making the IoT → cloud loop trustworthy enough that the analytics layer had clean event data to work from. Models are easy to demo; pipelines are what make models useful.