sm

Results — Model Accuracy

Sensor Evaluation

To assess the system’s performance, both the image classification model and the weight sensor were evaluated using 100 simulated waste samples across five material types. The summarized results are shown below:

Image Classification Model

The image classification model demonstrated strong overall performance, achieving 91% accuracy on a balanced test set of 100 items. It maintained consistent recall across all classes (approximately 90–95%), with clothes and metal showing the highest F1-scores of about 0.95. In contrast, plastic exhibited lower precision (0.78) due to visual similarity with white glass under varying lighting conditions. Overall, most misclassifications were limited to visually similar categories, indicating stable and reliable model behaviour.

Weight Sensor (Contamination Detection)

The weight sensor model achieved an overall accuracy of 88%, demonstrating strong reliability in detecting clean recyclables (precision ≈ 0.94). However, its performance for contaminated items was moderate, with a recall of about 0.72. These results suggest that while the system effectively identifies clean items, it occasionally misses or mis flags contamination, indicating room for improvement in detection accuracy.

Solution Limitation

a) Classification Accuracy (Camera)

The camera-based system employs a pre-trained 12-class AI model, but its accuracy can be affected by lighting conditions, object angles, and partially obscured items. The model may also misclassify unfamiliar local packaging and struggles when multiple items are deposited simultaneously. Moreover, classifying all plastics as recyclable and the limited 12-category scope reduce its ability to recognize complex or mixed waste items.

b) Residue Detection (Weight Sensor)

Residue detection uses a weight sensor and an estimated “empty weight” table, making it sensitive to variations in item size, packaging, or moisture. Over time, sensor drift can also cause measurement errors, leading to occasional false alerts or missed contamination. Since detection relies only on weight thresholds, it may not fully capture real-world cases such as oily plastics or wet paper.

c) User Behavior and Lack of Mechanical Enforcement

The prototype does not include motorized diverters, so users can ignore the tick/cross indicators and place items in the wrong bin. While education and signage can encourage proper disposal, the system ultimately relies on user cooperation, and some mis-sorting is likely to continue. Users may also repeatedly use the override option, marking recyclable items as non-recyclable, which could undermine the system’s accuracy.

Future Scope

a) Fine-tuning the image model with active learning:

The system can flag low-confidence images for human verification and retraining, gradually improving classification accuracy. Estimating item size helps refine expected weight ranges and reduce false residue alerts, while hardware acceleration (e.g., TensorRT or Coral TPU) can enhance real-time responsiveness.

b) Additional sensors:

Adding temperature and humidity sensors can help detect wet bins, while accelerometers may identify tampering or impacts. Barcodes or RFID tags on packaging could link items to expected weights and support simple recycling guidance.

c) Mechanical diverter gate:

A simple gate could automatically route items based on system confidence. When the Raspberry Pi is confident (e.g., ≥70%) and no residue is detected, the gate would direct the item to recycling; otherwise, it would remain or be redirected to general waste.

d) Multi-waste detection:

Instead of assuming one item per disposal, the bin could recognize multiple items at once. An object-detection model could identify several objects in a single frame, label each (e.g., plastic bottle, can, cardboard), and count them accurately.

e) Recyclable vs. non-recyclable differentiation:

The system could be upgraded to distinguish between recyclable and non-recyclable plastics using fine-tuned images, resin codes, or linked barcodes/RFID. A confidence-check mechanism could prompt user review when the model is unsure.

f) Low-power and outdoor scalability:

Future designs could use solar power to enable outdoor deployment without relying on mains electricity.

g) User-linked disposal and accountability:

Integrating RFID, card, or mobile login can track individual disposal behavior, enabling accountability, and supporting behavioral studies.

h) Predictive analytics for operations:

The system could forecast bin capacity and optimize collection routes, improving operational efficiency.

i) Bin-to-bin communication network:

For large-scale deployments, LoRa or NB-IoT connectivity would allow bins to communicate without relying on Wi-Fi infrastructure.

Conclusion

When we started this project, we weren’t just thinking about sensors and code, but about the overflowing bins, unpleasant odours, and recyclables mixed with general waste we all see every day. These problems led us to ask: Can technology make waste management smarter and more sustainable? Our IoT-powered smart bin represents a step toward that vision. By combining image classification, weight sensing, and fill-level monitoring, it not only detects when a bin is full but also understands what’s inside — reducing unnecessary collections and improving recycling quality. While sorting remains manual and the AI model needs fine-tuning for local packaging, the foundation is strong. With upgrades such as automated diverter gates and additional sensors, this solution can become even more effective. Ultimately, this prototype demonstrates how IoT can transform waste management from a reactive task into a proactive, data-driven strategy for cleaner and more sustainable communities.

References

https://www.imarcgroup.com/smart-waste-management-market

https://www.unep.org/resources/global-waste-management-outlook-2024

https://www.worldbank.org/en/topic/urbandevelopment/brief/solid-waste-management

https://recyclingpartnership.org/wp-content/uploads/dlm_uploads/2020/02/2020-State-of-Curbside-Recycling.pdf

https://nordsense.com/cases-san-francisco/

https://sensoneo.com/success-stories/prague-waste-collection-sensors/

https://controller.phila.gov/wp-content/uploads/2018/06/BigbellyReport_June2017.pdf