MarineCrewScenes: A Dataset for Multi-Scale Crew Monitoring in Ship Bridges

Navigating safely in the intricate operational environment of ship bridges can be significantly improved by reliable monitoring of crew behavior. However, applying existing object detection methods to monitor crew behavior confront many challenges including limited training data, complex backgrounds, fluctuating lighting conditions, and frequent crew movement. To address these challenges, we collected over 7,000 images to create a comprehensive, multi-scenario dataset, MarineCrewScenes, which encompasses diverse lighting conditions, crew postures, equipment interference, and varying perspectives, enhancing the model’s generalization capabilities. Based on this dataset, we propose an efficient crew detection model based on an improved RT-DETR, named Efficient Crew Detection Transformer (ECDETR). EC-DETR incorporates the Scale Sequential Feature Fusion (SSFF) module and Bi-Level Routing Attention (BLRA) mechanism to optimize detection across different scales, while integrating DIoU (Distance-IoU) loss for more precise bounding box regression. Experimental results demonstrate that EC-DETR achieves a high accuracy of 0.989 on the standard test set, and a significant improvement from 0.668 to 0.751 in average precision (AP50%) on the generalization test set. The model presents prominent generalization capability in unfamiliar environments that provides applicable solutions for crew monitoring in ship bridges and further maritime safety technologies.