An Integrated IoT and Edge-AI Framework for Real-Time Smart-City Waste Monitoring, Classification, and Collection-Route Optimization

Abstract

Urban waste collection still runs on fixed schedules that ignore how full bins actually are, so trucks arrive at empty bins (wasting fuel) and miss full ones (causing overflow, odour, and health risk). This paper presents an integrated IoT and edge-AI framework that closes that loop. Each smart bin pairs an ultrasonic sensor and a low-cost camera with an ESP32 microcontroller that estimates fill level on-device and publishes over Wi-Fi, GSM, or LoRaWAN to a cloud platform; the cloud predicts which bins will need collection and computes an optimized truck route. The framework is evaluated end to end. A median-filtered ultrasonic pipeline classifies bin state (empty/partial/full) at 92.0% accuracy, up from 85.9% raw. A from-scratch Random Forest predicts collection need at 89.7% accuracy (F1 0.897, ROC-AUC 0.938) over five runs, outperforming KNN by five points. Lightweight edge classifiers reach 84.5% on waste-type images at sub-millisecond inference - below deep transfer-learning models (89-94% from the literature) but small enough for on-device deployment. A nearest-neighbour + 2-opt route optimizer collects only full bins and cuts travel distance by 60.1% $(165 \rightarrow 65 \text{km})$, saving an estimated 34 L of fuel per cycle, and a duty-cycled node runs over 400 days on two 18650 cells. With hardware specifications, a dataset description, a scalability study to 1,000 bins, and a privacy design, the framework is a reproducible foundation for smart-city waste management.