Plug-and-Play Edge–Cloud Architecture for Autonomous Urban Garbage Detection with Self-Supervised Active Learning

Abstract

Urban waste accumulation in rapidly expanding Indian cities represents a persistent public health and governance crisis, compounded by the absence of scalable automated monitoring and the growing gap between waste generation and scientifically processed disposal. Existing AI-based waste detection systems are trained predominantly on Western datasets, require dedicated camera infrastructure, lack mechanisms for autonomous post-deployment improvement, and offer no accountability layer between detection and confirmed municipal action. This paper presents a plug-and-play dual-model edge-cloud waste intelligence framework that connects directly to existing street surveillance infrastructure without requiring any hardware replacement. A 2.9 MB YOLOv11n model runs real-time inference on resource-constrained edge devices — Raspberry Pi with under 500 MB RAM and no GPU — while a cloud-hosted verification model cross-validates detections, performs multi-class waste classification, and autonomously retrains the edge model weekly using cloud-corrected samples, eliminating continuous human annotation from the edge retraining loop entirely. The framework additionally counters illegal littering through timestamp-anchored detection logs, forecasts regional waste accumulation for proactive municipal planning, automates recycling coordination, and exposes a citizen-facing transparency platform that publicly reports the full lifecycle of each garbage event from detection to confirmed resolution. Initial experiments on TACO and TrashNet datasets yield mAP@0.5 of 45.8% and mAP@0.5:0.95 of 23.6% with an F1-optimised confidence threshold of 0.184, and real-time inference is confirmed on Raspberry Pi hardware under continuous operation without lag. The performance gap observed on Indian street conditions motivates an ongoing Pune-specific dataset collection effort whose domain adaptation results will be reported as the dataset matures. This architecture simultaneously addresses cost, scalability, domain specificity, legal deterrence, and governance transparency — making it directly deployable within Indian smart-city programmes at a fraction of the cost of purpose-built monitoring networks.