What it does
The smart bin leverages sensors, image processing, and machine learning to automatically identify and direct waste into the appropriate compartment. It also efficiently seals, replaces and disposes trash bags to avoid spills and easier maintenance.
Your inspiration
We were driven by the need to improve waste segregation in India, where only about 20% of municipal waste is recycled. Improper sorting at source leads to recyclable materials being contaminated and lost to landfills. The idea came from observing inefficient waste practices on campus and realizing that even educated communities struggle with proper segregation. Inspired by global innovations but aware of their limitations in Indian conditions, we envisioned a low-cost, sensor-based smart bin that could automate segregation accurately. Our goal was to design a practical, scalable solution tailored to local waste types.
How it works
When waste is thrown, a load cell weighs the item while a camera captures an image. Mixed waste is detected using weight and volume mismatch (from known density range) and diverted to general trash. Moisture sensors and Inductive Proximity sensors are used for liquid waste and metal respectively. This data is analyzed using our classification model. The model is trained on a custom dataset and uses VGG16 and SVM for classification with over 80% test accuracy. Based on sensor output and image classification, the lid rotates to direct the item into the correct compartment. A fast moving servo controlled flap. The inner bin made of wire mesh is partitioned into 6 sections using partition arms. These partition arms have cartridges of trash bags and heat sealers. When the bags are full, they are heat sealed and dropped where they can be collected through the bottom collection door. Bags are replaced by the cartridges in the arms, hence ease of use.
Design process
The design journey of the Smart Optimised Recycling Tool (S.O.R.T.) began in March 2024 with the aim of addressing poor waste segregation. After identifying limitations in current systems like high costs and low accuracy with mixed waste, the team focused on creating a low-cost, sensor-based, and AI-powered segregation bin. The first prototype was a conveyor belt system integrating an image classifier (85% accuracy using TrashNet) and moisture sensor. Though functional, it had limited precision and was not convenient for widespread use. We started with a hybrid feature-extraction pipeline (color, texture, deep learning features via VGG16) + SVM classifier. Later we added data augmentation , PCA for dimensionality reduction, hyperparameter tuning via GridSearchCV for getting upto 89% of accuracy. This led to a new rotating lid bin design with 6 compartments with a flap-controlled inlet. We used weight sensors, volume estimation, and material density comparison to flag mixed or contaminated items and divert them to general trash. To address confusion my ML model, we also introduced inductive proximity sensor to detect metal accurately. A heat-sealing mechanism was introduced to seal full bags automatically. An automatic bag replacement was also implemented for easy use.
How it is different
Unlike others, S.O.R.T. doesn’t assume clean waste. It checks for mismatches in weight, volume, and moisture to identify mixed or food-contaminated waste and reroutes it Most smart bins require manual emptying and replacing. S.O.R.T. features an automated heat-sealing mechanism that seals each compartment’s trash bag and drops it into a lower bin for easy collection. This reduces mess, labor, and risk of spills. Also, the setup will be disturbed less often making it more durable.
Future plans
We plan to pilot S.O.R.T. in IIT Bombay offices and campus hotspots to validate real-world performance. We’ll collect real-world waste data from campus deployments to retrain and fine-tune the models for Indian conditions. We also plan to include a post processing unit for plastic to this project (a different unit from the bin). This involves adding audio-based classification to distinguish materials like glass and plastic by their collision sounds and separating any metal contamination present through eddy currents. We’ll also integrate sensors ( using InGaAs photodiodes) to detect plastic types based on their spectral signatures.
Share this page on