Vision-based vessel detection for vessel-bridge collision warnings under complex scenes

Abstract

To enable accurate vessel recognition for bridge collision avoidance and early warning, an image dataset for vessels in bridge channels is established using cameras and data augmentation. This dataset includes complex scenarios such as long distances, multiple targets, and low visibility. Subsequently, the you-only-look-once version 5(YOLOv5)model is employed as the basic detector, and several modifications are applied to its network structure. Key enhancements involve replacing C3 modules in the backbone network with C2f modules, integrating the squeeze-excitation attention mechanism into the feature fusion network, and optimizing the prior anchors of the dataset using the K-means++ clustering algorithm. Finally, the modified model undergoes training and validation using PyTorch as the deep learning framework. Results demonstrate that the mean average precision for crucial vessels in the modified YOLOv5 model reaches 99.4%, representing an 11.1% improvement compared to the original YOLOv5 model. Additionally, the inference speed is measured at 102 frame/s. The established YOLOv5 model is a reliable and efficient cornerstone for warning against vessel-bridge collisions in complex navigable scenes.

Keywords

vessel detection / vessel-bridge collision / you-only-look-once version 5(YOLOv5) / squeeze-excitation attention mechanism / data augmentation