|
Development of an object detection model for detecting uncollected litter in Thailand |
|---|---|
| รหัสดีโอไอ | |
| Title | Development of an object detection model for detecting uncollected litter in Thailand |
| Creator | Suthat Daochern |
| Contributor | Dr. Prachya, Advisor |
| Publisher | Thammasat University |
| Publication Year | 2568 |
| Keyword | Object detection, YOLO, Uncollected litter, Computer vision, Waste management, Thailand uncollected litter dataset |
| Abstract | Thailand continues to face challenges related to municipal solid waste and uncollected litter, which affect environmental quality, public health, and the safety of sanitation workers. Although manual cleaning is commonly used to manage litter in public areas, this approach is limited in scale and may expose workers to hazardous working conditions, especially along roadsides and busy urban areas. Computer vision and deep learning offer a promising solution for supporting more proactive waste monitoring. However, existing litter detection datasets are mostly developed from foreign environments and may not fully represent the visual characteristics of litter and background conditions in Thailand.This study aimed to develop an object detection model for detecting uncollected litter in Thailand using YOLO-based deep learning models. A localized Thailand Uncollected Litter Dataset was developed, consisting of 2,112 images collected from real-world environments in Thailand, such as roadsides, local markets, public areas, canals, vacant lots, and coastal areas. The dataset was manually annotated using Roboflow, resulting in 2,940 annotated litter instances across 18 litter classes based on Thailand’s waste management context. The dataset was then divided into training, validation, and testing sets using an 80:10:10 ratio.Four YOLO models were trained and evaluated in this study, including YOLOv8n, YOLOv8m, YOLOv11n, and YOLOv11m. The model performance was assessed using mAP@0.5, precision, recall, and F1-score. The experimental results showed that YOLOv8m achieved the best overall performance, with an mAP@0.5 of 0.663 and an F1-score of 0.664. YOLOv11n achieved the highest precision of 0.783, while YOLOv11m achieved the highest recall of 0.633. These results indicate that different YOLO models have different strengths depending on the intended application. However, YOLOv8m provided the best balance between detection accuracy, localization quality, precision, and recall.The qualitative results showed that the model was able to detect several litter objects, such as plastic bottles, plastic bags, snack wrappers, plastic cups, straws, and other litter. Nevertheless, some prediction errors were still observed, including incorrect class prediction, incorrect localization, and missed detections, especially when objects were small, low contrast, visually similar, or located close to other objects. Overall, this study demonstrates that YOLO-based object detection models can be applied to detect uncollected litter in Thailand. The developed dataset and model can serve as a foundation for future waste monitoring systems, such as CCTV-based detection, drones, robots, or geographic information systems for supporting more efficient waste collection planning and improve waste management decision-making. |
