Urban Trash Detection and Cleanliness Assessment Based on YOLOv7 for Autonomous Sweeper

Urbanization accelerates, increasing municipal waste volumes. Automatic sweepers are crucial for urban cleanliness, but assessing ground cleanliness remains challenging, hindering efficient operations. We developed a ground cleanliness evaluation method based on urban waste recognition and detection. To mitigate the high false-negative rates and feature integration difficulties associated with diverse waste, we enhanced the YOLOv7 backbone with Res2Blocks for improved fine-grained feature fusion and detection accuracy. A quantifiable cleanliness index was formulated considering waste quantity, distribution, and area. A dataset of 10,543 images of typical urban park litter was compiled to train the model, achieving a mean Average Precision (mAP) of ≥0.926 at an Intersection over the Union (IoU) threshold of 0.5. This resulted in effective pre- and post-cleaning assessments, providing performance feedback and strategic guidance for future cleaning activities.