Multi-Feature Learning-Based Automatic Recognition of Non-Normative Seafarer Behaviours to Promote Maritime Traffic Safety

Abstract

In maritime navigation, seafarers' non-normative behaviours can significantly increase the likelihood of maritime accidents and lead to substantial losses. While monitoring equipment and computer vision technology are extensively employed in intelligent transportation systems (ITSs), behaviour detection within ship bridge situations is still rather scarce. We have constructed a dataset concentrating on non-normative behaviours within the ship's bridge environments, tackling the data scarcity problem in this domain. We initially extract essential information for later behaviour analysis by integrating an attention module with an object detection network, owing to the complexity of scenes in video surveillance. Meanwhile, we propose a behaviour recognition network utilizing multi-feature learning (termed MFLNet) to precisely assess seafarer activities in critical areas. In particular, MFLNet adaptively synthesizes seafarer appearance and posture through a compression and incentive module, enhancing recognition accuracy and mitigating sample imbalance issues. Extensive qualitative and quantitative experiments indicate that the MFLNet attains superior speed and accuracy for recognizing non-normative seafarer behaviours.