Vehicle Trajectory Prediction at Roundabouts Based on Time Series Pattern Decomposition

To enhance the vehicle trajectory prediction accuracy in complex structured scenarios such as roundabouts, a deep learning model combining multi-head attention (MHA) mechanism, simplified graph convolution (SGC) network and TimeMixer mechanism, namely MST, is proposed. The model is built upon a macro-micro dual-encoder architecture. At the macro level, an MHA mechanism is employed to capture the long-term guiding constraints imposed by global road topology. That is, by modeling the complete historical trajectory of the vehicle and the structure of the roundabout (such as the deep relationship of the entrances and exits) to infer vehicle’s long-term driving intention. At the micro level, first, an SGC network is used to extract instantaneous spatial relationships among vehicles. Subsequently, TimeMixer mechanism is introduced to map the one-dimension interaction sequence into multi-scale, multi-resolution 2D spatio-temporal images. By explicitly decoupling and hierarchically fusing periodic tactical behaviors and trend-oriented strategic intentions, a precise capture of deep interaction patterns is achieved. The information streams from both levels are integrated via a gated fusion network and then fed into a gated recurrent unit decoder to generate the final trajectory. Experiments on the public INTERACTION and RounD datasets demonstrate that, within a 5 s prediction period, the proposed model achieves an average displacement error and a final displacement error of 1.19 m and 1.85 m on the INTERACTION dataset, and 1.16 m and 1.80 m on the RounD dataset, respectively, outperforming all baseline models. The results indicate that hierarchically modeling macro-level global constraints and micro-level spatio-temporal interactions, particularly through the decoupling analysis of interaction patterns, can significantly improve the trajectory prediction performance in complex scenarios.