[Paper Review] LAformer: Trajectory Prediction for Autonomous Driving with Lane-Aware Scene Constraints

Mengmeng Liu, Hao Cheng+6 more|arXiv (Cornell University)|Feb 27, 2023

Autonomous Vehicle Technology and Safety6 citations

TL;DR

LAformer proposes a two-stage trajectory prediction framework for autonomous driving that uses a temporally dense lane-aware module to dynamically select top-k lane segments at each future time step, aligning motion dynamics with scene constraints. It further improves accuracy via a second-stage motion refinement module, achieving state-of-the-art performance on Argoverse 1 and nuScenes benchmarks with 0.64 ADE₆ and 0.92 FDE₆ on Argoverse 1.

ABSTRACT

Trajectory prediction for autonomous driving must continuously reason the motion stochasticity of road agents and comply with scene constraints. Existing methods typically rely on one-stage trajectory prediction models, which condition future trajectories on observed trajectories combined with fused scene information. However, they often struggle with complex scene constraints, such as those encountered at intersections. To this end, we present a novel method, called LAformer. It uses a temporally dense lane-aware estimation module to select only the top highly potential lane segments in an HD map, which effectively and continuously aligns motion dynamics with scene information, reducing the representation requirements for the subsequent attention-based decoder by filtering out irrelevant lane segments. Additionally, unlike one-stage prediction models, LAformer utilizes predictions from the first stage as anchor trajectories and adds a second-stage motion refinement module to further explore temporal consistency across the complete time horizon. Extensive experiments on Argoverse 1 and nuScenes demonstrate that LAformer achieves excellent performance for multimodal trajectory prediction.

Motivation & Objective

• To address the challenge of inaccurate trajectory predictions in complex scenes, such as intersections, where existing one-stage models fail to maintain consistent alignment between motion dynamics and HD map constraints.
• To reduce representation burden on the decoder by filtering irrelevant lane segments through a dynamic, time-aware lane selection mechanism.
• To improve long-horizon temporal consistency by introducing a second-stage motion refinement module that corrects predictions from the first stage.
• To enable more accurate and scene-compliant multimodal trajectory forecasts by leveraging vectorized HD maps and attention-based feature learning.

Proposed method

• A Global Interaction Graph (GIG) encoder processes unified vectorized trajectories and HD map data to extract joint spatial-temporal features.
• A binary classifier, trained on GIG-embedded features and agent motion states (speed, orientation), estimates the likelihood of each lane segment aligning with motion dynamics at each time step.
• The top-k most likely lane segments are selected per time step to condition the subsequent attention-based decoder, reducing irrelevant map noise.
• A Laplacian Mixture Density Network (MDN) generates multimodal trajectory predictions aligned only with the selected lane segments.
• A second-stage motion refinement module takes observed and first-stage predicted trajectories as input to refine predictions and reduce temporal offsets.
• The model is trained end-to-end with multi-task loss: a classification loss for lane selection and a prediction loss for trajectory generation.

Experimental results

Research questions

• RQ1 Can a temporally dense lane-aware selection mechanism improve trajectory prediction accuracy by continuously aligning motion dynamics with scene constraints?
• RQ2 Does decoupling lane selection from trajectory prediction via a two-stage framework enhance long-horizon temporal consistency?
• RQ3 How does filtering irrelevant lane segments affect the decoder’s representation efficiency and prediction quality?
• RQ4 To what extent does a second-stage motion refinement module reduce prediction errors compared to one-stage baselines?

Key findings

• LAformer achieves 0.64 ADE₆ and 0.92 FDE₆ on the Argoverse 1 benchmark, outperforming prior state-of-the-art models.
• On nuScenes, LAformer achieves 0.65 ADE₆ and 0.95 FDE₆, demonstrating strong generalization across diverse traffic scenarios.
• Ablation studies show that removing the temporally dense lane-aware module reduces prediction diversity and accuracy, especially in turning and acceleration scenarios.
• The second-stage refinement module significantly improves prediction quality, with the complete model outperforming ablated variants in both quantitative and qualitative evaluations.
• The model maintains robustness across varying hyperparameters, with minimal performance fluctuation under different loss weight settings.
• LAformer’s inference speed of ~115ms per scenario with 12 agents is comparable to LTP and faster than DenseTNT, making it suitable for near real-time deployment.