Map-Free Trajectory Prediction in Traffic with Motion-Pattern-Enhanced Hyper-Relational Interaction Modeling

Trajectory prediction in complex traffic scenes requires both accurate and real-time. Most of the existing advanced map-based Trajectory prediction methods rely heavily on high-definition maps for higher accuracy, but the high-definition maps in open driving environments are extremely expensive to construct and maintain, which hinders the generalization of such models. Another type of map-less methods have a faster reasoning speed, but they perform much lower in complex scenes than the map-based methods due to the lack of map information constraints. To this end, we present a two-stage map-less Trajectory prediction method. First, a contrastive learning-based encoder captures motion patterns, providing implicit constraint semantic. Second, an adaptive multi-granularity hypergraph topology construction module and a bi-directed hypergraph attention module capture complex interaction patterns among agents. Extensive experiments demonstrate that our model achieves state-of-the-art performance when comparing with advanced map-less methods and competitive performance against most map-based methods.