A Hybrid YOLO–Swin Transformer with FSM for Waste Disposal Behavior Analysis

Abstract

The increasing volume of waste and improper disposal practices remains significant challenges in smart campus environments despite the availability of waste management facilities. In this study, waste disposal behavior is defined as the temporal and contextual interaction among individuals, waste objects, and trash bins, encompassing both proper disposal and littering activities. To address this problem, this research proposes a hybrid deep learning framework that integrates YOLOv8 for real-time object detection, a Finite State Machine (FSM) for modeling object-state transitions during disposal activities, and a Swin Transformer for contextual behavior analysis. The proposed framework extends conventional object detection by combining spatial detection, temporal interaction modeling, and contextual reasoning to classify waste disposal behaviors from surveillance data. Furthermore, the integration of YOLOv8, FSM, and Swin Transformer provides a unified and interpretable architecture capable of capturing object interactions and behavioral patterns in complex surveillance environments. Experimental results show that the proposed method achieves a behavior classification mAP@50 of 93.8%, outperforming the FSM-only baseline (90.0%) and improving precision, recall, and F1-score from 0.89, 0.84, and 0.86 to 0.92, 0.88, and 0.90, respectively. These findings demonstrate that the combination of spatial, temporal, and contextual analysis enables more accurate and reliable waste disposal behavior classification, providing a practical solution for intelligent waste monitoring and sustainable smart campus management.