Intention-fused multi-model trajectory prediction for lane-changing

Trajectory prediction is a crucial foundation for safe decision-making in autonomous driving systems, particularly in strongly interactive scenarios such as lane changing. To address the semantic dilution of driving intention in traditional shallow feature fusion and the limited stability of single-model prediction across different temporal horizons, this paper proposes a multi-model trajectory prediction framework integrating intention priors and temporal weighting mechanisms. Specifically, driving intention is introduced as a high-level semantic prior into an attention-enhanced LSTM architecture to strengthen the modeling of critical trajectory points; simultaneously, an LSTM-based fusion network is designed to perform temporal weighting on the predictions of heterogeneous models, effectively combining their respective advantages to generate the optimal trajectory. Experimental results demonstrate that the proposed multi-model fusion significantly improves prediction accuracy and robustness. Compared to single-model benchmarks, the Average Displacement Error on the NGSIM and HighD datasets is reduced by 12.20%–24.56% and 14.96%–22.96%, respectively; more importantly, the Lateral Prediction Error is reduced by 24.59%–50.13% on the NGSIM dataset and 18.01%–51.05% on the HighD dataset. These results confirm that the synergy of multi-model modeling effectively enhances full-horizon prediction stability and ensures that predicted trajectories are highly consistent with real-world driving behavior.