ABSTRACT
This study developed an automatic monitoring system for Floating Marine Debris (FMD) aimed at reducing the labor-intensiveness of traditional visual surveys. It involved creating a comprehensive FMD database using 55.6 h of video footage and numerous annotated images, which facilitated the training of a deep learning model based on the YOLOv8 architecture. Additionally, the study implemented the BoT-SORT algorithm for FMD tracking, significantly enhancing detection accuracy by effectively filtering out disturbances such as sea waves and seabirds, based on the movement patterns observed in FMD trajectories. Tested across 16 voyages in various marine environments, the system demonstrated high accuracy in recognizing different types of FMD, achieving a mean Average Precision (mAP@0.5) of 0.97. In terms of detecting FMD from video footage, the system reached an F1 score of 83.63 %. It showed potential as a viable substitute for manual methods for FMD larger than 20 cm.