A YOLO Benchmarking Experiment For Maritime Object Detection In Foggy Environments

The maritime object detection in foggy conditions has been a major challenge in the field of computer vision due to its harsh environment. Compared to automatic detection in other domains, progress in this field has been slow, primarily due to the lack of benchmarking studies. In this study, we propose a benchmark study for maritime obj ect detection in foggy environments by combining the modified ABOships dataset and the efficient YOLO (You Only Look Once) model. We simulate foggy conditions using the principle of dark channel prior and modify the original dataset through class balancing and data augmentation techniques. Furthermore, through training and analysis, we select the state-of-the-art YOLOv5n model as the base model. The experimental results are encouraging, as the YOLO model trained using modified dataset demonstrates a notable improvement in mAPO.5 by 15.82% and mAPO.5:0.95 by 16.6%, respectively. Moreover, the proposed base model demonstrates superior stability compared to other models, achieving an mAPO.5 of 64.4% while requiring only 3.64 million parameters and 4.3 GFLOPs (Giga Floating Point Operations Per Second) of computational cost. This effectively demonstrates that the proposed methodology can be used as a reproducible basis for comparative evaluation in foggy maritime object detection, thereby advancing research in smart transportation services.