Computer vision / 2023 — 2024

Smart Waste Classification

A vision model connected to a physical sorting loop.

Drishti / Makernova · Hardware prototype Source code

EDGE AI / RASPBERRY PIRECORDED PROTOTYPE

Original prototype recording from the public repository. No live inference. Open video file

Visual description: a plastic bottle rests on a white platform inside a cardboard enclosure, with a handheld camera assembly above it. At the end of the demonstration, the platform is clear. The model and hardware sequence are explained below.

01 / THE PROBLEM

Intelligence at the edge.

A classifier becomes a different engineering problem when it must react to a sensor and control physical hardware. This prototype closes that loop for waste separation.

My contribution

AI/ML project contributor

I developed the classifier, collected images and integrated the model with a Raspberry Pi. The prototype connects camera capture and prediction to indicator LEDs and a servo-driven sorting mechanism.

02 / TECHNICAL APPROACH

Inside the
implementation.

01
An IR sensor triggers a webcam capture. The image is resized to 64 × 64, converted to an array and normalized by 255.
02
A saved VGG16 model predicts a class. The hardware path maps the outcome into biodegradable and non-biodegradable servo positions.
03
The Raspberry Pi drives GPIO LEDs and PWM servo movement, then returns the actuator to its neutral position.

03 / SYSTEM ARCHITECTURE

Follow the signal.

Select a stage to explore
how the system works.

IR sensor

A GPIO input detects a nearby object and triggers capture.

IR sensor: A GPIO input detects a nearby object and triggers capture.

Camera: OpenCV captures a frame. Keras preprocessing resizes to 64 × 64 and normalizes pixel values.

VGG16 model: The hardware script loads a saved classifier and selects the argmax output.

Servo sorting: The binary hardware route sets one of two PWM positions, then resets to neutral.

05 / RESULTS & OUTCOMES

What the work delivers.

2Physical sorting outcomes

64 × 64Classifier input resolution

The prototype connects an IR trigger, camera capture, VGG16 classification and servo actuation in one physical loop. Its hardware route separates biodegradable and non-biodegradable waste and resets the servo after each cycle.

Scope & limitations

The video is an existing prototype recording, not live model inference.
Model files, hardware setup and runtime dependencies are required to reproduce the experiment.

06 / ENGINEERING DECISIONS

The interesting
constraints.

Connecting model and mechanism

The implementation coordinates an input sensor, camera, model, status LEDs and physical actuation. Safe cleanup stops PWM and resets GPIO on interruption.

ENGINEERING TAKEAWAY

On-device AI includes the sensor and actuator lifecycle, not only model inference.

07 / BUILDING BLOCKS

Tools with
a purpose.

TensorFlow / Keras: Image classification
OpenCV: Webcam capture
Raspberry Pi / GPIO: Sensor and actuator control

Go to the source

Hardware source Repository and recordings

NEXT PROJECT / 06