LA-GAT boosts AV safety with lane-aware trajectory prediction in merge zones

Accurate trajectory prediction in merge zones is critical for autonomous vehicle safety, yet existing graph-based models are designed for mainline freeways and ignore geometrically distinct merge interactions. Researchers propose the Lane-Aware Graph Attention Network (LA-GAT), which encodes vehicle interactions within dynamic scene graphs augmented with a trainable lane-relationship attention bias. This bias prioritizes merge-conflict vehicles from the start of training, improving prediction in complex merge and diverge areas. The model is pre-trained on raw NGSIM US-101 and I-80 datasets, then fine-tuned on UAV-captured UTE SQM-W-1 trajectory data from a Chinese expressway merge area.

Evaluation on the held-out SQM-W-2 dataset shows strong performance: ADE of 0.865m at 1s and 2.518m at 3s, demonstrating that drone-informed fine-tuning substantially reduces cross-dataset transfer gaps. The model also incorporates surrogate safety metrics beyond standard displacement errors—including TTC violation rate, DRAC exceedance rate, and collision rate. Deliberate use of unfiltered NGSIM data reveals raw-condition generalization limits due to measurement errors, providing valuable insights for real-world deployment. LA-GAT directly enables safer autonomous driving decisions in highway merge zones.