Relative Motion-Aware and Multi-Scale Trend Fusion Network for Long-Term Trajectory Prediction
2026-99-2451
11/20/2026
- Content
- Accurately predicting the future trajectories of surrounding traffic participants is a critical foundation for collision avoidance and local-path planning in autonomous driving systems, as reliable motion prediction directly affects the safety and robustness of downstream decision-making modules. However, existing prediction methods still face significant challenges in long-term trajectory prediction over a horizon of 3-5 s. On the one hand, interaction modeling methods based on Euclidean distance struggle to effectively capture non-local inter-vehicle interaction over a large spatial range. On the other hand, single-scale temporal encoders are susceptible to local trajectory fluctuations and short-term noise, making it difficult to extract the stable long-term motion trend and limiting their ability to model complex temporal dependency. To address these issues, this paper proposes a Relative Motion-aware and Multi-scale Trend Fusion Network (RM-Net) for long-term trajectory prediction. First, a Relative Motion-aware Non-local Interaction Modeling Mechanism (RNM) is designed to overcome the limitation of traditional spatial proximity criteria. Guided by the relative distance change rate between vehicles, RNM leverages a cross-attention mechanism to accurately model distant yet high-risk non-local interaction feature. Second, a Multi-scale Motion Trend Adaptive Fusion module (MTM) is constructed to extract and adaptively fuse vehicle motion trends under different temporal resolutions, enabling more robust sequential modeling of the trajectory. Extensive ablation experiments on the NGSIM dataset validate the effectiveness and complementary benefits of the proposed RNM and MTM modules. Furthermore, comparative experiments against several state-of-the-art trajectory prediction methods demonstrate that the proposed RM-Net achieves significantly superior long-term prediction accuracy compared with existing state-of-the-art models.
- Citation
- Chen, S., Zhang, X., Tian, L., Zhang, G., et al., "Relative Motion-Aware and Multi-Scale Trend Fusion Network for Long-Term Trajectory Prediction," 2026 2nd International Conference on Intelligent Transportation and Future Mobility (ITFM2026), Xi'An, China, March 27, 2026, .