Abstract
As a crucial part of social behavior modeling, multi-agent trajectory prediction with dynamic interactions has become one of the core technologies for autonomous systems. One of the main research streams is to leverage relational graph Transformers to learn interactions among agents, as this method demonstrates superior capability in capturing implicit features and modeling complex relationships. However, such researches fall short in efficiently utilizing multi-domain features, consequently being constrained by spatial attention layers that hinder accuracy improvements. To address the above limitations and achieve a more lightweight map-free architecture, this work introduces a Multi-domain Hybrid Graph Transformer (MHTraj) for predicting multi-modal future trajectories. The core insight of the method lies in its well-designed encoder, which enhances inputs via a frequency-domain mixer and employs causal convolution attention to capture global temporal information. Furthermore, a graph Transformer is incorporated to aggregate spatial relationships, thereby yielding the final hybrid features. Extensive experiments on 4 real-world datasets (NBA, ETH/UCY, SDD, nuScenes) show that the proposed method achieves competitive state-of-the-art (SOTA) performances, and meanwhile reducing the parameter count by nearly 50% on the ETH/UCY datasets.
Funding Information
  • National Science and Technology Innovation 2030 (2022ZD0205005)
  • National Natural Science Foundation of China (61573057)