Yolo-Based Multi-Model Ensemble for Plastic Waste Detection Along Railway Lines

A rapidly increasing amount of plastic waste not only cause serious environmental issues but also pose a considerable threat to the rail transportation. It is important to monitor the intrusion of floating plastics into the railway area. In this article, we propose to detect plastic waste using You Only Look Once-v5 (YOLO-v5) algorithm and model ensemble through surveillance cameras installed along railway lines. Experiments on the size of YOLO-v5 model were carried out to find the optimal size to detect plastics. The model with large size (YOLOv51) outperformed with an overall accuracy (OA) of 82.6% and mean Average Precision (mAP) of 0.822. Two ensemble modelling strategies were implemented considering different size combination of YOLO-v5 models including 1) nano, small and medium sizes; 2) nano, small, medium and large sizes. The latter one achieved the best result with the OA equal to 85.4% and the mAP equal to 0.834. The results indicate that YOLO-based ensemble model can effectively improve the performance of detection plastic waste using surveillance cameras and the acquired knowledge has great potential to UAV- and satellite-based high-resolution imagery.