Human-in-the-Loop Enhanced YOLOv8 Intelligent Waste Bin for Real-Time Waste Classification

Abstract

Waste management remains a significant environmental challenge in urban areas, where inadequate waste segregation reduces recycling efficiency and increases pollution. Although computer vision enables automated waste classification, many existing systems rely on static datasets and cannot adapt to changing conditions after deployment. This study proposes an intelligent waste bin integrating YOLOv8n with a Human-in-the-Loop (HITL) feedback framework to support real-time classification and iterative model improvement. The prototype combines a Raspberry Pi 5, Arduino Nano, USB camera, graphical user interface, and servo-based sorting mechanism to identify and separate paper, plastic, and organic waste. A dataset of 965 annotated images from campus environments and public sources was used for training and evaluation. MobileNetV2, EfficientNetB0, ResNet50, and YOLOv8n were compared in offline and real-world deployment tests. YOLOv8n showed the most robust deployment performance, achieving 75.00% accuracy, 85.19% precision, 74.44% recall, and 73.16% F1-score. HITL-based user corrections and retraining further improved classification performance. At the 10% HITL stage, offline evaluation yielded 91.19% accuracy, 91.44% precision, 91.23% recall, and 91.33% F1-score. Following HITL integration, deployment evaluation achieved 100% across these metrics on the tested samples. These findings suggest that incorporating human feedback can enhance the adaptability and practical performance of embedded waste classification systems, although broader deployment testing is needed to establish long-term reliability.