Abstract
Multi-agent trajectory prediction plays a crucial role in various domains, including autonomous driving, unmanned systems, and robotics. However, existing approaches have not fully addressed the complex interplay between motion patterns and multi-agent interaction features, making accurate trajectory prediction in complex scenarios particularly challenging. To address these limitations, we present Nomoformer, a transformer-based architecture for multi-agent trajectory prediction. Nomoformer predicts agent trajectories through joint analysis of motion representations and spatio-temporal interactions. Our approach first models individual agent dynamics by incorporating physical constraints derived from velocity and angular features. We mapped select motion features to the frequency domain to capture inherent relationships between periodic patterns and agent topology. To model inter-agent interactions, we developed a spatio-temporal graph topology using heterogeneous graphs with directed edge features and attention masks. Within this topology, we introduced Edge Encoder and Node Encoder modules that effectively establish long-range dependencies across edge features and temporal dimensions. Extensive experiments across multiple benchmarks demonstrated that Nomoformer achieved state-of-the-art performance in predicting future motion trajectories.
