CoFUSE-EPWNet: Cross-Modal Consistency Fusion for RGB-D Express Packaging Waste Detection in Green Campus Management

Area of Science:

• Robotics and Automation
• Computer Vision
• Waste Management Engineering

Background:

• Rapid growth of express packaging waste (EPW) poses significant urban waste management challenges.
• Automated sorting of EPW is hindered by material heterogeneity, deformability, stacking, and occlusion.
• Current sorting systems lack the reliability needed for high-quality resource recovery from EPW.

Purpose of the Study:

• To develop a robust RGB-D sensing framework for reliable automated sorting of express packaging waste (EPW).
• To create a comprehensive dataset (MEPWaste) for training and evaluating EPW sorting systems.
• To demonstrate the engineering feasibility of an integrated sensing-inference-execution pipeline for EPW recycling.

Main Methods:

• Construction of the MEPWaste dataset: 4528 synchronized RGB-D samples from 10 packaging categories.
• Design of a multimodal detection framework leveraging RGB-D data for enhanced feature reliability and spatial consistency.
• Implementation of a closed-loop pipeline: synchronized acquisition, real-time detection, robotic grasping, and PLC control.

Main Results:

• The proposed RGB-D framework achieved 89.1% mAP50 and 69.3% mAP50:95, outperforming RGB-only and baseline RGB-D models.
• Significant improvements in detection accuracy (5.3-5.8 pp over RGB-only, 2.7-3.4 pp over baseline) with 85.1% recall.
• Successful demonstration of a closed-loop system for practical engineering feasibility in EPW sorting.

Conclusions:

• The developed RGB-D sensing framework significantly enhances the reliability of automated EPW sorting.
• The study provides a viable technical solution for intelligent resource recovery in waste management.
• The system shows potential for application in high-consumption environments like green campuses.