A TCN trajectory prediction method fusing MSC and spatio-temporal dual attention

Aiming at the problem that existing trajectory prediction models are difficult to effectively extract multi-scale spatio-temporal features, which leads to limited prediction accuracy, a new method MDAT-Net for trajectory prediction based on a temporal convolutional network (TCN) fused with a multi-scale convolutional (MSC) network and spatio-temporal dual attention (STDA) is proposed. The MDAT-Net model contains the MSAT, MTAT trajectory prediction module and voting module. First, in the prediction module, a multi-scale convolution architecture is built using different scale convolution kernels to better extract spatio-temporal features at various scales and solve the fixed kernel size issue in the traditional temporal convolutional network. Secondly, in order to dynamically mine the potential correlation between hidden features and target features, spatial attention mechanism and temporal attention mechanism are introduced to adaptively focus on important information and skip secondary information. Finally, the voting module decides which model to apply for each dimension prediction, allowing the benefits of both prediction models to be combined and high-precision trajectory prediction to be achieved. The experimental results show that the MDAT-Net model can improve the root mean square error (RMSE) up to 83.33% and the mean absolute error (MAE) up to 85.85% with high accuracy and robustness.