Intent-aware contrastive learning for trajectory prediction under varying observation lengths

Trajectory prediction for autonomous driving is challenged by inconsistent observation lengths, which induce distribution shifts in trajectory encoder representations and degrade performance under varying observation lengths. The existing method aligns trajectory representations across different observation lengths via contrastive learning. However, it does not explicitly model high-level behavioral semantic similarity among trajectories, potentially introducing false-negative supervision. To address this issue, we propose an intent-aware contrastive learning framework that encourages representation consistency across varying observation lengths while promoting high-level intent semantic consistency in the representation space. We introduce an adaptive negative debiasing mechanism that continuously modulates the weights of semantically consistent and representation-similar negative samples, thereby alleviating erroneous supervision in representation learning. The proposed framework improves the stability and robustness of the trajectory encoder under varying observation lengths. Experimental results on benchmark datasets show that the proposed method consistently achieves performance gains under varying observation lengths, outperforming the existing method.