Leveraging Scene-Invariant Priors for Autoregressive Human Trajectory Prediction

Abstract

Accurate human trajectory prediction is essential for autonomous driving and robot navigation. Despite substantial progress, deep learning-based approaches often suffer from the train-inference gap caused by distributional discrepancies between training and test environments. The goal-guided framework mitigates this issue by exploiting the scene-invariant prior that human motion is typically goal-driven. However, existing goal-guided approaches do not fully leverage additional scene-invariant priors and therefore still face two key limitations: insufficient diversity in predicted goals and the lack of explicit modeling of human kinematic consistency. In this paper, we propose SIPTraj, a goal-guided trajectory prediction framework that integrates scene-invariant priors at both the long-term intention and short-term motion levels. For goal estimation, we introduce goal candidates derived from offline clustering of trajectory endpoints as a data-driven approximation of scene-invariant long-term intention modalities, and develop a weighted Farthest Point Sampling (FPS) strategy that balances spatial coverage with predicted likelihoods to improve goal diversity while preserving semantic plausibility. For trajectory completion, we develop a kinematic-aware dual-stream autoregressive decoder that jointly predicts aligned position and velocity. The decoder adopts a kinematic initialization with residual refinement (KIRR) strategy: future states are initialized using a constant-velocity kinematic model and refined via reciprocal cross-attention, which injects kinematic consistency into the decoding process and leads to more natural and accurate trajectory prediction. Extensive experiments demonstrate that SIPTraj achieves state-of-the-art performance on the ETH/UCY and SDD benchmarks.