Timeba: UNet State Space Model for Trajectory Prediction

In recent years, the State Space Model (SSM)-based sequence modeling technique, specifically Mamba, has achieved remarkable success and widespread utilization in both Natural Language Processing and Computer Vision. Mamba can effectively capture long-range dependencies of sequence but lacks the ability to address multi-scale local sequence information. By combining Mamba's strength in modeling long-range dependencies with UNet's ability to extract multi-level local details, we explore for the first time extending state space model-based methods to trajectory prediction, and we design an effective temporal trajectory prediction model named Timeba using a U-shaped architecture. Experiments were executed on three real-world traffic datasets, yielding highly competitive accuracy for both short-term and long-term prediction tasks. In addition, an empirical analysis was conducted concerning the contribution of diverse trajectory parameters, to determine their effect on the final prediction precision.