Citation

Ding H, Weng J, Han B. J. Transp. Saf. Secur. 2023; 15(12): 1299-1324.

Copyright

(Copyright © 2023, Southeastern Transportation Center, and Beijing Jiaotong University, Publisher Informa - Taylor and Francis Group)

DOI

10.1080/19439962.2023.2169801

PMID

unavailable

Abstract

The accurate detection of seafarers' unsafe behaviors is of great significance to improve the ship navigation safety. This study proposes an improved deep learning framework with a self-made dataset to detect the unsafe behavior of seafarers on duty. In order to increase the detection speed, the improved Cross Stage Partial connections (CSP) module is proposed to replace the original CSP module in the neck network of traditional algorithm. The efficient channel attention (ECA) module is also introduced to the backbone network of conventional algorithm as the attention mechanism network. In addition, the learning and representation capacities of the improved deep learning framework are promoted by redesigning the sizes of anchor boxes. The experiment results show that the proposed framework outperforms traditional object detection algorithms (e.g., YOLOv5s, R-CNN) in detecting seafarers' unsafe behaviors because of the much higher detection speed and detection accuracy.

Language: en

Keywords

human factor; marine traffic; object detection; tracking control; unsafe behavior