Edge-enriched Graph Transformer for Multi-Agent Trajectory Prediction with Relative Positional Semantics

Trajectory prediction is critical for safe and efficient autonomous driving, especially in scenarios with intricate road structures and complex interactions. To address this challenge, we propose a framework based on edge-enriched graph transformers for multi-modal trajectory prediction of multiple agents. The model is novel in interaction representation and unified input format. First, to model the interaction, an edge-featured graph is constructed with relative coordinates, and positional semantics as edge properties, where the position information like front, rear, left, right, and conflict modes are encoded using binary codes. The edge features capturing the interaction relationship are further used for closeness recognition. Second, we achieve a unified representation of map and agent features, ensuring the consistent scale and interpretation of the heterogeneous input. Specifically, we vectorize and discretize the lanes into agent-like units, and individualise the lanelets with the agent-specific features. To handle the graph-like input, the edge-enriched graph transformer is first introduced for feature encoding. Finally, the dynamic, interaction, and map features are concatenated for multi-modal prediction decoding. The experiments are conducted using the INTERACTION dataset and Argoverse2. The results of the comparison and ablation experiments demonstrate the competitive performance of our model in the highly interactive scenes compared with other state-of-the-art prediction methods.