Litter detection in Real-World Environments using YOLO models

The automatic detection of litter in real-world environments is crucial for advancing smart waste management systems and promoting environmental sustainability. Real-time garbage identification can help mitigate the negative impacts of litter on ecosystems and urban areas. In this work, we utilize the pLitterStreet dataset, which provides two detection schemes: a 4-class scheme (Plastic, Pile, Face mask, and Trash bin) and a 10-class scheme (Litter, Pile, Face mask, Trash bin, Plastic bag, Bottle, Cup, Rope, Sachet, and Straw). For both experiments, we use multiple YOLO models, including YOLOv8l, YOLOv9c, and YOLOv10l. Among these, for the 4-class experiment, YOLOv8l achieved the highest mAP50 score of 72% and 51% of mAP, outperforming the baseline results from the original pLitterStreet dataset introduction paper. Whereas for the 10-class detection task, our best model was YOLOv10l which achieved 40.0% of mAP50 and 24.4% of mAP. Finally, we used the best-performing model on a 4-class scheme (YOLOv8l) to detect the presence of litter in the streets of eight major cities worldwide. We used the Google Street View API to obtain images in a 10 km ratio from each city center and generated heat maps to visualize the areas with litter present.