DCSTNet: A Dual-Channel Spatio-Temporal Information Fusion Network for Map-Free Vehicle Trajectory Prediction

The complex interaction between traffic participants brings safety problems for autonomous driving in mixed-traffic environment. Current state-of-the-art (SOTA) vehicle trajectory prediction models suffer significant performance degradation when high-definition (HD) map inputs are excluded, which may compromise the safety of decision-making and planning for autonomous driving systems in real-world traffic environments. To address this challenge, this paper proposes a novel dual-channel interactive modelling framework, termed the DCSTNet (dual-channel spatio-temporal information fusion network), specifically designed for vehicle trajectory prediction without relying on HD map information. Unlike previous trajectory prediction models that model temporal and spatial interactions interlacing or hierarchically, DCSTNet decoupling temporal and spatial interaction modules through a specially designed encoding network. This practice enables the model to more fully extract interaction features without increasing computational complexity when map information is not considered. To verify the validity of the dual-channel spatio-temporal information fusion framework, our study uses the publicly available Argoverse motion forecasting dataset. The comparison of results demonstrates that DCSTNet outperforms many SOTA approaches, including those that use map-based priors. To further validate that decoupling temporal and spatial interaction modelling enhances feature extraction capabilities, we conduct rigorous ablation studies and sensitivity analysis on the dataset to dissect architectural components of the DCST network. To explore the adaptability of the framework, we also develop a map-based variant of DCSTNet and compare its predictions with the map-free version in complex road environments.