基于二维卷积和Transformer的船舶轨迹预测方法OA

This paper proposes a 2D-CNN-Transformer model for ship trajectory prediction,integrating 2D convolution neu-ral networks with Transformer architecture to address the limitations of discrete position encoding and single modal feature extraction in conventional methods.The model employs continuous position encoding to eliminate quantization errors induced by grid-based co-ordinate mapping,reconstructs trajectory images for spatial feature extraction via ResNet,and leverages cross-attention for dynam-ic multi-modal feature fusion.Experimental results demonstrate superior performance in mean square error(MSE)and root mean square error(RMSE)across 5 min,10 min,and 15 min prediction horizons compared to baseline models.The synergistic integra-tion of continuous position encoding,two-dimensional convolutional branches,and cross-attention mechanisms offers a novel para-digm for high-precision trajectory prediction within intelligent shipping systems.