Urban Waste Detection and Segmentation using Deep Learning and Computer Vision Techniques

In this paper, a deep learning-based urban solid waste detection and segmentation approach based on street-level images is proposed. The work targets the challenge of waste monitoring in intricate urban environments by leveraging a mix of object detection and segmentation techniques. This framework introduces a prompt-based interaction layer that bridges the YOLOv8-based waste detection model and the Segment Anything Model (SAM2). A specialized dataset, Dataset A, with 9231 images and 9490 annotations, was created to simulate real urban waste scenarios. The system is tested on Dataset A and on two publicly available datasets, UAV-USW and pLitterStreet. Model A outperformed all other approaches on Dataset A with an mAP50 of 0.659 and precision and recall of 0.73 and 0.59, respectively. The performance in cross-dataset evaluation exhibited substantial differences, due to the varying types of domains and annotations. The results show that training on realistic datasets greatly enhances the detection of waste in urban areas. The proposed framework serves as a basis to develop a smart waste monitoring system for city administration.