Abstract:

To enhance vehicle trajectory prediction accuracy in complex structured scenarios such as roundabouts, a deep learning framework, namely MST (MHA-SGC-TimeMixer), is proposed. The framework is built upon a macro-micro dual-encoder architecture. At the macro level, a Multi-Head Attention (MHA) mechanism is employed to capture the long-term guiding constraints between vehicles and the global road topology. At the micro level, a Simplified Graph Convolutional Network (SGC) first extracts instantaneous spatial relationships among vehicles. Subsequently, the TimeMixer mechanism is introduced to map the one-dimensional interaction sequence into multi-scale, multi-resolution 2D spatio-temporal images. By explicitly decoupling and hierarchically fusing periodic tactical behaviors and trending 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 (GRU) decoder to generate the final trajectory. Experiments on the public INTERACTION and RounD datasets demonstrate the framework's effectiveness. Within a 5-second prediction horizon, the proposed model achieves an Average Displacement Error (ADE) and a Final Displacement Error (FDE) of 1.19m and 1.85m on the INTERACTION dataset, and 1.16m and 1.80m 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 trajectory prediction performance in complex scenarios.

Key words:

roundabout, trajectory prediction, multi-head attention, simplified graph convolution, TimeMixer