INTraJ: A Unified Framework for Trajectory Prediction with Explicit Planning and Reaction Decomposition
A recent publication on arXiv (2608.05673) presents INTraJ, a comprehensive framework designed for trajectory prediction that separates social influence into two distinct phases: planning and reaction. During the planning phase, reference trajectories are generated based on anticipated social dynamics, while the reaction phase focuses on making local adjustments derived from the difference between full-context predictions and these references. This framework accommodates both single-target and multi-target scenarios. Testing across four established benchmarks—Argoverse 2, Argoverse 2-ped, ETH/UCY, and SDD—reveals significant enhancements, especially in managing social interactions. This new submission has been authored by a team of researchers.
Key facts
- INTraJ is a unified framework for trajectory prediction.
- It decomposes social influence into planning and reaction stages.
- The planning stage uses future social information to construct reference trajectories.
- The reaction stage recovers local adjustments from residuals.
- INTraJ supports both multi-target and single-target paradigms.
- Experiments were conducted on Argoverse 2, Argoverse 2-ped, ETH/UCY, and SDD.
- Results show consistent improvements, especially in social interaction handling.
- The paper is available on arXiv with ID 2608.05673.
Entities
Institutions
- arXiv