Intelligent shipping systems advance deep learning applications in ship trajectory prediction, enhancing maritime security risk identification. However, existing models struggle to capture multi-ship interactions in high-density areas due to the significant influence of neighboring vessels. To address this, we propose a trajectory prediction method integrating kullback-leibler divergence attention (KLDA) and a dual-stream parallel architecture. First, a transposition interactive convolutional network (TICNet) employs a dual-layer convolution structure to model dynamic interactions between target and adjacent ships. Second, KLDA mitigates amplitude interference by mapping feature vectors to probability distributions, decoupling similarity computation from input amplitude to suppress dominant local outliers in attention maps. Third, the time-feature dual-stream strategy separately models temporal dependencies via time-embedding sequences and encodes heterogeneous features as tokens using sample transposition, with self-attention explicitly capturing feature correlations. Finally, a cascaded transformer decoder (CTD) iteratively refines outputs through recursive optimization. Experimental results on three real-world scenarios demonstrate average displacement error reductions of 16.75%, 33.48%, and 45.63% compared to the best baseline, validating the method’s effectiveness in complex, high-density navigation environments.
原文链接
分享
引用
