MR BIN is a smart, automated waste management system developed as a senior thesis project at Ateneo de Zamboanga University (ADZU). The device is designed to identify, measure, and compress PET bottles while rejecting non-plastic materials, providing a reward system to encourage sustainable recycling habits.
- Intelligent Detection: Distinguishes PET bottles from other materials using Machine Learning.
- Precision Measurement: Calculates bottle volume via Computer Vision.
- Volume Reduction: Mechanically compresses bottles to maximize internal storage efficiency.
- Incentive System: Provides user rewards ("Incentive Noting") upon successful transactions.
The device is divided into three core functional modules:
- Plastic Detection Module: Utilizes a camera and an ML model to verify object types.
- Volume Measurement Module: Employs image processing to calculate physical dimensions.
- PET Bottle Storage Module: Manages the mechanical compression and storage workflow.
Initially, the system utilized four ultrasonic distance sensors within a 40cm x 20cm x 20cm enclosure. The goal was to calculate dimensions based on the distance between the sensors and the bottle:
-
Diameter (
$d$ ):$X - (x_1 + x_2)$ -
Height (
$h$ ):$Y - (y_1 + y_2)$ -
Volume Formula:
$V = \pi \left(\frac{d}{2}\right)^2 h$
The Challenge: Despite integrating temperature sensors to calibrate the speed of sound, the approach proved unreliable. Because the radius is squared in the volume equation, even microscopic "noise" in the ultrasonic readings caused exponential errors in the final volume calculation.
To achieve higher accuracy, we transitioned to OpenCV. The system now captures a top-down image, detects edges, and creates a mask to encapsulate the bottle within a bounding rectangle.
- Dimensioning: We calculate the Euclidean distance between the four bounding points.
- Calibration: Using a Pixel-per-Metric constant, the system converts pixel distances into real-world metric units. This is calibrated based on the fixed height of the camera from the base.
For robust object recognition, we utilized TensorFlow and a Transfer Learning approach.
- Model:
ssd-mobilenet-lite(selected for its high efficiency on Raspberry Pi hardware). - Performance: After extensive training, the model achieved an average loss of 0.05. We carefully tuned the training to avoid overfitting, ensuring the Raspberry Pi can handle real-time inference during the recycling process.
- Hardware: Raspberry Pi, Pi Camera, Mechanical Actuators.
- Software/Libraries: TensorFlow, OpenCV, Python.
- Model Architecture: SSD-MobileNet-Lite.
Developed as a Senior Thesis Project at Ateneo de Zamboanga University (ADZU).