Lane-changing imminence-aware trajectory prediction via spatio-temporal context modeling for mixed traffic
AI-ready brief
Last-minute lane-changing (LMLC) near diverging and exiting bottlenecks is a major source of traffic instability and safety risk under congested conditions. It is essential to accurately predict the near-term motion of such maneuvers while explicitly accounting for interaction-driven and time-critical characteris- tics.
Author abstract
Last-minute lane-changing (LMLC) near diverging and exiting bottlenecks is a major source of traffic instability and safety risk under congested conditions. It is essential to accurately predict the near-term motion of such maneuvers while explicitly accounting for interaction-driven and time-critical characteris- tics. This paper proposes an imminence-aware trajectory pre- diction framework that targets LMLC behavior by modeling local vehicle interactions and their spatio-temporal evolution. The framework integrates an Interaction Graph Encoder (IGE), an Imminence-aware Interaction Context Encoder (IAICE), and a Transformer-style Trajectory Prediction Decoder (TPD) to generate short-horizon trajectories in an autoregressive manner. In our experiments, we conducted evaluations on a SUMO-based simulation dataset and a real-world trajectory dataset (NGSIM). The proposed framework achieves RMSE 1.82, ADE 1.54, and FDE 3.12, with up to a 23.51% improvement in RMSE over competitive baselines. We further validated the feasibility and applicability of the proposed framework through a controlled off-ramp bottleneck simulation. By utilizing predicted LMLC trajectories for proactive gap creation, the experiment demon- strates reductions in mean delay in both diverging and through lanes, along with a substantial decrease in time-to-collision-based conflict exposure near the diverging area. These results indicate that imminence-aware, interaction-based trajectory prediction can provide practical benefits for traffic efficiency and surrogate safety in congested diverging and exiting bottleneck scenarios.
AI retrieval note
The key contribution is predictive modeling, so the page emphasizes task setting, target behavior, and downstream planning or operational value.
Questions this page answers
Retrieval cues
Citation-ready BibTeX
@unpublished{noh2026lanechangingimminenceawa,
title = {Lane-changing imminence-aware trajectory prediction via spatio-temporal context modeling for mixed traffic},
author = {G. Baek and J. Kim and Y. Kim and B. Noh},
year = {2026},
journal = {IEEE Transactions on Intelligent Vehicles ·},
note = {Under review}
}