Lane-Aware Graph Attention Network for Multi-Vehicle Trajectory Prediction in Expressway Merge Zones

Technical Summary: Lane-Aware Graph Attention Network for Multi-Vehicle Trajectory Prediction in Expressway Merge Zones

Problem Statement
Accurate multi-vehicle trajectory prediction is critical for autonomous decision-making, particularly in expressway merge and diverge zones where interaction structures are geometrically distinct and complex. Existing graph-based prediction models are predominantly developed and validated on mainline freeway segments using fixed-camera datasets like NGSIM. These models often fail to address the specific interaction topologies of merge zones. Furthermore, standard evaluation protocols rely exclusively on displacement error metrics (ADE, FDE), leaving the safety consequences of predicted trajectories unquantified. A significant gap exists in the literature regarding the generalization of models trained on data with known measurement errors (such as NGSIM) to new environments, and the lack of surrogate safety measure (SSM) evaluation in trajectory prediction.

Methodology
The paper proposes a Lane-Aware Graph Attention Network (LA-GAT) designed to encode vehicle interactions within dynamic scene graphs. The methodology involves four key components:

1. Dynamic Scene Graph Construction: At each time step, the traffic scene is represented as a directed graph where nodes are vehicles and edges represent influence. Edges are established based on:
   • Proximity: A threshold derived from the 95th percentile of observed inter-vehicle distances.
   • Structural Constraints: Explicit leader-follower records and mandatory edges for vehicles in adjacent lanes within a ±10 m longitudinal separation to ensure merge-conflict pairs are always represented.
2. Lane-Aware Attention Mechanism: The model utilizes a trainable lane-relationship attention bias. Before attention computation, the lane relationship feature (mapping to same, left, right, or merging) is modified by a learnable scalar vector. This mechanism is initialized to assign a positive bias specifically to "merging" relationships, structurally prioritizing merge-conflict interactions from the onset of training.
3. Architecture:
   • Temporal Encoder: A bidirectional LSTM processes 30-frame (3-second) historical state sequences (position, speed, acceleration, normalized lane ID, lane-change flag) to produce 128-dimensional motion context vectors.
   • Graph Attention: Two stacked GATv2 layers aggregate neighborhood information using the lane-bias-modified edge features.
   • Decoder: Three independent MLP heads predict bivariate Gaussian parameters for 1s, 3s, and 5s horizons.
4. Cross-Dataset Transfer Protocol:
   • Pre-training: The model is pre-trained on raw, unfiltered NGSIM US-101 and I-80 datasets. No quality filtering is applied to leader-follower records or headway values to benchmark performance under real-world data quality conditions.
   • Fine-tuning: The BiLSTM encoder is frozen, while the GATv2 layers and decoder are fine-tuned on UAV-captured UTE SQM-W-1 trajectory data from a Chinese expressway merge area.
   • Evaluation: The model is evaluated on the held-out UTE SQM-W-2 dataset (zero-shot and fine-tuned conditions).

Key Contributions

1. LA-GAT Model: Introduction of a graph attention network with a trainable lane-relationship bias that explicitly prioritizes merge-conflict interactions during training.
2. Cross-Dataset Transfer Protocol: A novel workflow involving pre-training on raw NGSIM data and fine-tuning on high-resolution UAV data (UTE SQM), marking the first application of Chinese expressway drone-BEV data to GAT-based trajectory prediction.
3. Raw-Data Generalization Analysis: Deliberate use of unfiltered NGSIM data to characterize the generalization limits and performance degradation caused by measurement errors (e.g., invalid leader IDs), directly addressing data quality sensitivity issues identified in prior benchmarks.
4. Expanded Evaluation Metrics: Extension of evaluation beyond displacement errors to include Surrogate Safety Measures (SSM): Time-to-Collision (TTC) violation rates, Deceleration Rate to Avoid Collision (DRAC) exceedance rates, and collision rates computed from reconstructed absolute predicted positions.

Results

• Displacement Accuracy: Fine-tuning on UAV data significantly reduced the cross-dataset transfer gap. On the held-out SQM-W-2 dataset, the fine-tuned model achieved an Average Displacement Error (ADE) of 0.865 m at 1s and 2.518 m at 3s. This represents a reduction of approximately 79% and 77% respectively compared to the zero-shot NGSIM-only baseline (4.125 m and 10.779 m).
• NGSIM Performance: On the NGSIM test set, the model exhibited high error rates (ADE of 17.39 m at 1s) and high SSM violation rates. The authors attribute this to the use of raw, unfiltered data containing corrupted interaction records, confirming that measurement errors in training data directly limit model generalization.
• Safety Metrics: The fine-tuned model demonstrated lower TTC violation and DRAC exceedance rates at the 1s horizon compared to the zero-shot baseline, suggesting that proximity to actual vehicle positions reduces spurious near-miss detections. However, high violation rates persisted at longer horizons (e.g., TTC violation rising from 56.74% at 1s to 84.84% at 5s), reflecting the inherent uncertainty of human gap-acceptance behavior in merge zones rather than solely model error.

Significance and Claims
The paper claims that drone-informed fine-tuning is essential for bridging the gap between models trained on legacy, error-prone datasets and real-world merge zone applications. The study demonstrates that while pre-training on raw NGSIM data provides a baseline, the structural differences and measurement errors in that dataset prevent effective zero-shot transfer to Chinese expressway geometries.

The authors emphasize that the rapid degradation of prediction accuracy over time (from 0.865 m at 1s to 5.034 m at 5s) and the elevated safety violation rates at longer horizons are not merely model limitations but reflect the compounding uncertainty inherent in human gap-acceptance behavior at merge points. The study posits that once a merging vehicle commits to a gap, its trajectory becomes heavily conditioned on non-deterministic reactions from mainline vehicles, fundamentally bounding the predictability horizon in these zones.

Limitations Acknowledged by the Authors

• The TTC safety penalty term was set to zero during training due to computational costs ($O(n^2)$) on CPU hardware.
• No ablation studies (e.g., constant velocity, LSTM-only, Social LSTM) were conducted to isolate the specific contribution of the graph attention mechanism and lane-aware bias.
• The study did not apply quality filtering to the NGSIM pre-training data to create a "cleaned" baseline, though this was a deliberate choice to test raw-condition generalization.