pith. sign in

arxiv: 2605.11940 · v2 · submitted 2026-05-12 · 📡 eess.SY · cs.SY

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

Eni Solomon Laughter This is my paper

Pith reviewed 2026-05-14 20:59 UTC · model grok-4.3

classification 📡 eess.SY cs.SY
keywords trajectory predictiongraph attention networksmerge zoneslane-aware attentionUAV fine-tuningsurrogate safety metricsautonomous drivingdynamic scene graphs
0
0 comments X

The pith

A lane-aware graph attention network with bias for merge conflicts predicts trajectories more accurately after drone-data fine-tuning.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes that adding a trainable lane-relationship attention bias to graph attention networks lets the model prioritize merge-conflict interactions from the start of training. Pre-training occurs on raw NGSIM freeway segments and fine-tuning uses UAV-captured trajectories from a Chinese expressway merge, with evaluation on held-out data from the same site. This matters because autonomous vehicles need reliable short-horizon forecasts in geometrically distinct merge zones to support safe lane-change decisions, and the work also tracks safety consequences through time-to-collision and deceleration-rate metrics rather than displacement error alone. Results show the fine-tuned model reaches 0.865 m ADE at the one-second horizon on the target merge dataset, narrowing the gap that appears when models trained only on mainline data are applied directly.

Core claim

The Lane-Aware Graph Attention Network encodes multi-vehicle interactions inside dynamic scene graphs and augments them with a trainable lane-relationship attention bias that prioritizes merge-conflict pairs. Pre-training on unfiltered NGSIM US-101 and I-80 data followed by fine-tuning on UAV SQM-W-1 trajectories yields an ADE of 0.865 m at 1 s and 2.518 m at 3 s on the held-out SQM-W-2 set, while also reporting TTC violation rates and DRAC exceedance rates to quantify safety implications.

What carries the argument

Trainable lane-relationship attention bias inside the Lane-Aware Graph Attention Network (LA-GAT), which adjusts attention weights on dynamic scene graphs to emphasize interactions between vehicles on merging lanes.

If this is right

  • Displacement errors drop measurably when the model is exposed to merge-specific geometry during fine-tuning.
  • Surrogate safety indicators such as TTC violation rate become quantifiable outputs alongside trajectory forecasts.
  • Unfiltered NGSIM data exposes the raw generalization ceiling imposed by sensor noise in legacy datasets.
  • Domain adaptation via UAV footage bridges the distribution shift between mainline freeways and merge zones.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same explicit lane-bias mechanism could be tested on intersections or weaving sections where lane conflicts also dominate.
  • Collecting UAV trajectories from additional merge geometries would allow direct checks on whether the fine-tuning benefit holds across countries or road standards.
  • If lane geometry is encoded as an attention bias rather than a hard constraint, the model may adapt more easily to temporary lane closures or construction zones.
  • The reported safety-metric improvements suggest the framework could be coupled with risk-aware planning modules that penalize high-TTC-violation forecasts.

Load-bearing premise

The lane-relationship attention bias reliably learns to focus on merge-conflict interactions and the UAV fine-tuning data forms a representative sample of the target merge-zone distribution without new biases absent from the NGSIM source.

What would settle it

If removing the lane-relationship attention bias or skipping the UAV fine-tuning step produces equal or lower ADE and safety-metric violation rates on SQM-W-2, the claimed benefit of the proposed components would be refuted.

read the original abstract

Accurate multi-vehicle trajectory prediction in expressway merge and diverge areas is fundamental to the decision-making frameworks of autonomous vehicle systems. However, the majority of existing graph-based prediction models are developed and validated on mainline freeway segments and do not address the geometrically distinct interaction structures that characterize merge zones. Furthermore, standard evaluation protocols rely exclusively on displacement error metrics, leaving the safety consequences of predicted trajectories unquantified. This paper proposes a Lane-Aware Graph Attention Network (LA-GAT) that encodes vehicle interaction within dynamic scene graphs, augmented with a trainable lane-relationship attention bias that prioritizes merge-conflict interactions from the outset of training. The model is pre-trained on the raw NGSIM US-101 and I-80 datasets and subsequently fine-tuned on UAV-captured UTE SQM-W-1 trajectory data from a Chinese expressway merge area, with final evaluation on the held-out SQM-W-2 dataset. Evaluation spans both displacement metrics (ADE, FDE at 1s, 3s, 5s horizons) and surrogate safety measures (TTC violation rate, DRAC exceedance rate, collision rate). Fine-tuned results on SQM-W-2 yield ADE of 0.865 m at 1s and 2.518 m at 3s, demonstrating that drone-informed fine-tuning substantially reduces the cross-dataset transfer gap. The deliberate use of unfiltered NGSIM data is shown to characterize raw-condition generalization limits, with the performance degradation attributed to the well-documented measurement errors in that dataset.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript introduces the Lane-Aware Graph Attention Network (LA-GAT) for multi-vehicle trajectory prediction in expressway merge zones. It encodes interactions via dynamic scene graphs augmented by a trainable lane-relationship attention bias, pre-trains on raw NGSIM US-101 and I-80 data, fine-tunes on UAV-captured UTE SQM-W-1 trajectories, and evaluates on held-out SQM-W-2 using ADE/FDE at 1s/3s/5s horizons plus surrogate safety metrics (TTC violation rate, DRAC exceedance rate, collision rate). The central claim is that drone-informed fine-tuning substantially reduces the cross-dataset transfer gap, with reported fine-tuned ADE of 0.865 m at 1 s and 2.518 m at 3 s on SQM-W-2.

Significance. If the lane-aware bias demonstrably prioritizes merge-conflict edges and the fine-tuning distribution is representative, the work would provide a concrete advance in handling geometrically distinct merge-zone interactions that standard graph models overlook. The explicit use of unfiltered NGSIM data to expose raw-condition generalization limits is a methodological strength that aligns with realistic deployment constraints.

major comments (3)
  1. [Abstract] Abstract: the claim that the trainable lane-relationship attention bias 'prioritizes merge-conflict interactions from the outset of training' is not supported by any ablation (e.g., bias-ablated baseline) or attention-weight visualization on merge scenes; the reported ADE/FDE gains on SQM-W-2 could arise entirely from fine-tuning on UAV data rather than the bias mechanism.
  2. [Evaluation] Evaluation protocol: no comparison of pre-trained versus fine-tuned attention maps is shown to verify that the bias term up-weights merge-related edges rather than generic proximity; without this isolation, the central assertion that the bias reduces the transfer gap remains unverified.
  3. [Safety metrics] Safety metrics section: surrogate quantities (TTC violation rate, DRAC exceedance rate) are reported without any quantified correlation to actual collision events in the datasets, so the claim that the model improves safety consequences rests on untested surrogates.
minor comments (2)
  1. [Abstract] The abstract states that architecture diagrams, loss formulations, and ablation studies are absent; including them would improve reproducibility and allow readers to assess the exact form of the lane-relationship bias.
  2. [Introduction] The specific measurement errors in NGSIM that are invoked to explain performance degradation should be referenced to the original NGSIM documentation or prior quantification studies.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments, which help clarify the presentation of our contributions. We address each major point below and indicate the revisions we will make to the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim that the trainable lane-relationship attention bias 'prioritizes merge-conflict interactions from the outset of training' is not supported by any ablation (e.g., bias-ablated baseline) or attention-weight visualization on merge scenes; the reported ADE/FDE gains on SQM-W-2 could arise entirely from fine-tuning on UAV data rather than the bias mechanism.

    Authors: We agree that the abstract claim requires explicit supporting evidence. In the revised manuscript we will add an ablation study that removes the lane-relationship attention bias while keeping all other components fixed, thereby isolating its contribution to the reported ADE/FDE improvements on SQM-W-2. We will also include attention-weight visualizations on representative merge scenes from the held-out SQM-W-2 data to show that the bias term preferentially weights merge-conflict edges rather than generic proximity. These additions will demonstrate that the performance gains are not attributable solely to fine-tuning. revision: yes

  2. Referee: [Evaluation] Evaluation protocol: no comparison of pre-trained versus fine-tuned attention maps is shown to verify that the bias term up-weights merge-related edges rather than generic proximity; without this isolation, the central assertion that the bias reduces the transfer gap remains unverified.

    Authors: We accept that a direct pre-trained versus fine-tuned comparison is necessary to substantiate the mechanism. The revised manuscript will contain a new subsection and accompanying figure that extracts and contrasts attention maps from the NGSIM-pre-trained model and the UTE-fine-tuned model on identical merge-zone scenes. This will illustrate how the lane-aware bias modulates edge weights specifically for merge-conflict interactions and thereby narrows the cross-dataset transfer gap. revision: yes

  3. Referee: [Safety metrics] Safety metrics section: surrogate quantities (TTC violation rate, DRAC exceedance rate) are reported without any quantified correlation to actual collision events in the datasets, so the claim that the model improves safety consequences rests on untested surrogates.

    Authors: The NGSIM and UTE datasets consist of naturalistic trajectories that contain no actual collision events; therefore a direct empirical correlation between the surrogate metrics and observed collisions cannot be computed within these data. TTC and DRAC are nevertheless standard, literature-validated proxies for collision risk in trajectory-prediction studies. In the revision we will expand the safety-metrics discussion to cite the relevant validation studies and to clarify that the reported reductions in violation rates are presented as improvements in surrogate safety indicators rather than as direct collision predictions. revision: partial

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper describes an architecture (LA-GAT with trainable lane-relationship attention bias) pre-trained on NGSIM then fine-tuned on SQM-W-1 and evaluated on held-out SQM-W-2. No equations, derivations, or self-citations are shown that reduce the reported ADE/FDE or safety metrics to fitted parameters by construction. The performance numbers are obtained via standard train/fine-tune/evaluate protocol on distinct datasets; no self-definitional, fitted-input-called-prediction, or ansatz-smuggling patterns appear. The central claim rests on empirical results rather than a closed logical loop.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The central claim rests on the effectiveness of the lane-relationship attention bias and successful domain transfer from NGSIM to UAV data; no explicit free parameters, axioms, or invented entities are stated in the abstract.

pith-pipeline@v0.9.0 · 5578 in / 1141 out tokens · 39787 ms · 2026-05-14T20:59:15.934769+00:00 · methodology

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

26 extracted references · 26 canonical work pages

  1. [1]

    Trajectory data-based traffic flow studies: A revisit,

    L. Li, R. Jiang, Z. He, X. (Michael) Chen, and X. Zhou, “Trajectory data-based traffic flow studies: A revisit,” TRANSPORTATION RESEARCH PART C-EMERGING TECHNOLOGIES, vol. 114. PERGAMON-ELSEVIER SCIENCE LTD, THE BOULEVARD, LANGFORD LANE, KIDLINGTON, OXFORD OX5 1GB, ENGLAND, pp. 225–240, May

    work page

  2. [2]

    doi: 10.1016/j.trc.2020.02.016

    work page doi:10.1016/j.trc.2020.02.016 2020

  3. [3]

    Teaching cars to drive,

    F. N. Tzortzoglou and A. A. Malikopoulos, “Teaching cars to drive,” IEEE Potentials, vol. 44, no. 6, pp. 15–24, Nov. 2025, doi: 10.1109/MPOT.2026.3664382

    work page doi:10.1109/mpot.2026.3664382 2025

  4. [4]

    A review of vehicle lane change research,

    C. Ma and D. Li, “A review of vehicle lane change research,” Phys. Stat. Mech. Its Appl., vol. 626, p. 129060, Sep. 2023, doi: 10.1016/j.physa.2023.129060

    work page doi:10.1016/j.physa.2023.129060 2023

  5. [6]

    Merging control strategies of connected and autonomous vehicles at freeway on-ramps: A comprehensive review,

    J. Zhu, S. Easa, and K. Gao, “Merging control strategies of connected and autonomous vehicles at freeway on-ramps: A comprehensive review,” J. Intell. Connect. Veh., vol. 5, no. 2, pp. 99–111, May 2022, doi: 10.1108/JICV-02-2022- 0005

    work page doi:10.1108/jicv-02-2022- 2022

  6. [7]

    The characteristics of driver lane-changing behaviour in congested road environments,

    W. Wang and G. Cheng, “The characteristics of driver lane-changing behaviour in congested road environments,” TRANSPORTATION SAFETY AND ENVIRONMENT, vol. 6, no. 3. OXFORD UNIV PRESS, GREAT CLARENDON ST, OXFORD OX2 6DP, ENGLAND, Jul. 02, 2024. doi: 10.1093/tse/tdad039

    work page doi:10.1093/tse/tdad039 2024

  7. [8]

    Modeling Lane-Changing Behavior in Freeway Off-Ramp Areas from the Shanghai Naturalistic Driving Study,

    L. Zhang, C. Chen, J. Zhang, S. Fang, J. You, and J. Guo, “Modeling Lane-Changing Behavior in Freeway Off-Ramp Areas from the Shanghai Naturalistic Driving Study,” JOURNAL OF ADVANCED TRANSPORTATION. WILEY- HINDAWI, ADAM HOUSE, 3RD FL, 1 FITZROY SQ, LONDON, WIT 5HE, ENGLAND, 2018. doi: 10.1155/2018/8645709

    work page doi:10.1155/2018/8645709 2018

  8. [9]

    A Car-Following Model Considering Preceding Vehicle’s Lane-Changing Process,

    M. Zhao, S.-H. Wang, D. Sun, and X.-J. Wang, “A Car-Following Model Considering Preceding Vehicle’s Lane-Changing Process,” IEEE ACCESS, vol. 7. IEEE-INST ELECTRICAL ELECTRONICS ENGINEERS INC, 445 HOES LANE, PISCATAWAY, NJ 08855-4141 USA, pp. 89913– 89923, 2019. doi: 10.1109/ACCESS.2019.2924659

    work page doi:10.1109/access.2019.2924659 2019

  9. [10]

    A data-driven lane-changing model based on deep learning,

    D. Xie, Z. Fang, B. Jia, and Z. He, “A data-driven lane-changing model based on deep learning,” Transp. Res. PART C-Emerg. Technol., vol. 106, pp. 41–60, Sep. 2019, doi: 10.1016/j.trc.2019.07.002

    work page doi:10.1016/j.trc.2019.07.002 2019

  10. [11]

    Comparison of Machine Learning Algorithms for Predicting Lane Changing Intent,

    D. Choi and S. Lee, “Comparison of Machine Learning Algorithms for Predicting Lane Changing Intent,” Int. J. Automot. Technol., vol. 22, no. 2, pp. 507–518, Apr. 2021, doi: 10.1007/s12239-021-0047-x

    work page doi:10.1007/s12239-021-0047-x 2021

  11. [12]

    Detecting lane change maneuvers using SHRP2 naturalistic driving data: A comparative study machine learning techniques,

    A. Das, M. N. Khan, and M. M. Ahmed, “Detecting lane change maneuvers using SHRP2 naturalistic driving data: A comparative study machine learning techniques,” ACCIDENT ANALYSIS AND PREVENTION, vol. 142. PERGAMON-ELSEVIER SCIENCE LTD, THE BOULEVARD, LANGFORD LANE, KIDLINGTON, OXFORD OX5 1GB, ENGLAND, Jul. 2020. doi: 10.1016/j.aap.2020.105578

    work page doi:10.1016/j.aap.2020.105578 2020

  12. [13]

    A Driving-Preference-Aware Framework for Vehicle Lane Change Prediction,

    Y. Lyu, Y. Wang, H. Liu, X. Dong, Y. He, and Y. Ren, “A Driving-Preference-Aware Framework for Vehicle Lane Change Prediction,” SENSORS, vol. 25, no. 17, Aug. 2025, doi: 10.3390/s25175342

    work page doi:10.3390/s25175342 2025

  13. [14]

    Detection of Driver Styles in Lane Changes using Wavelet Transform,

    Y. E. Avci and A. Tuncer, “Detection of Driver Styles in Lane Changes using Wavelet Transform,” JOURNAL OF UNIVERSAL COMPUTER SCIENCE, vol. 30, no. 5. GRAZ UNIV TECHNOLGOY, INST INFORMATION SYSTEMS COMPUTER MEDIA-IICM, INFFELDGASSE 16C, GRAZ, A-8010, AUSTRIA, pp. 590–602, 2024. doi: 10.3897/jucs.108073

    work page doi:10.3897/jucs.108073 2024

  14. [15]

    A combined use of microscopic traffic simulation and extreme value methods for traffic safety evaluation,

    C. Wang, C. Xu, J. Xia, Z. Qian, and L. Lu, “A combined use of microscopic traffic simulation and extreme value methods for traffic safety evaluation,” Transp. Res. PART C-Emerg. Technol., vol. 90, pp. 281–291, May 2018, doi: 10.1016/j.trc.2018.03.011

    work page doi:10.1016/j.trc.2018.03.011 2018

  15. [16]

    A full Bayes approach for traffic conflict-based before-after safety evaluation using extreme value theory,

    L. Zheng and T. Sayed, “A full Bayes approach for traffic conflict-based before-after safety evaluation using extreme value theory,” Accid. Anal. Prev., vol. 131, pp. 308–315, Oct. 2019, doi: 10.1016/j.aap.2019.07.014

    work page doi:10.1016/j.aap.2019.07.014 2019

  16. [17]

    FollowNet: A Comprehensive Benchmark for Car-Following Behavior Modeling,

    X. Chen et al., “FollowNet: A Comprehensive Benchmark for Car-Following Behavior Modeling,” Sci. DATA, vol. 10, no. 1, Nov. 2023, doi: 10.1038/s41597-023-02718-7

    work page doi:10.1038/s41597-023-02718-7 2023

  17. [18]

    Trajectory data reconstruction and simulation-based validation against macroscopic traffic patterns,

    M. Montanino and V. Punzo, “Trajectory data reconstruction and simulation-based validation against macroscopic traffic patterns,” Transp. Res. PART B-Methodol., vol. 80, pp. 82–106, Oct. 2015, doi: 10.1016/j.trb.2015.06.010

    work page doi:10.1016/j.trb.2015.06.010 2015

  18. [19]

    A Modified Car- following Model Considering Traffic Density and Acceleration of Leading Vehicle,

    X. Cao, J. Wang, and C. Chen, “A Modified Car- following Model Considering Traffic Density and Acceleration of Leading Vehicle,” Appl. Sci.- BASEL, vol. 10, no. 4, Feb. 2020, doi: 10.3390/app10041268

    work page doi:10.3390/app10041268 2020

  19. [20]

    A review of car-following and lane- changing models under heterogeneous environments,

    Y. Chen, C. Dong, K. Lyu, X. Shi, G. Han, and H. Wang, “A review of car-following and lane- changing models under heterogeneous environments,” Phys. -Stat. Mech. ITS Appl., vol. 654, Nov. 2024, doi: 10.1016/j.physa.2024.130127

    work page doi:10.1016/j.physa.2024.130127 2024

  20. [21]

    A hybrid physics-based and data-driven approach for car-following behavior modeling and analysis,

    Q. Cheng et al., “A hybrid physics-based and data-driven approach for car-following behavior modeling and analysis,” Transp. Res. Part C Emerg. Technol., vol. 177, p. 105207, Aug. 2025, doi: 10.1016/j.trc.2025.105207

    work page doi:10.1016/j.trc.2025.105207 2025

  21. [22]

    Adaptive physics-informed trajectory reconstruction exploiting driver behavior and car dynamics,

    M. A. Makridis and A. Kouvelas, “Adaptive physics-informed trajectory reconstruction exploiting driver behavior and car dynamics,” Sci. Rep., vol. 13, no. 1, p. 1121, Jan. 2023, doi: 10.1038/s41598-023-28202-1

    work page doi:10.1038/s41598-023-28202-1 2023

  22. [23]

    A perspective from competitive-cooperative driving modes: Identification of vehicle merging behavior models and crash risk factors in merge zone,

    N. Lyu, S. Feng, and H. Wan, “A perspective from competitive-cooperative driving modes: Identification of vehicle merging behavior models and crash risk factors in merge zone,” Accid. Anal. Prev., vol. 220, Sep. 2025, doi: 10.1016/j.aap.2025.108153

    work page doi:10.1016/j.aap.2025.108153 2025

  23. [24]

    A Macroscopic Model for Freeway Weaving Sections,

    F. Marczak, L. Leclercq, and C. Buisson, “A Macroscopic Model for Freeway Weaving Sections,” COMPUTER-AIDED CIVIL AND INFRASTRUCTURE ENGINEERING, vol. 30, no. 6, SI. WILEY, 111 RIVER ST, HOBOKEN 07030- 5774, NJ USA, pp. 464–477, Jun. 2015. doi: 10.1111/mice.12119

    work page doi:10.1111/mice.12119 2015

  24. [25]

    Application of machine learning algorithms in lane-changing model for intelligent vehicles exiting to off-ramp,

    C. Dong, H. Wang, Y. Li, X. Shi, D. Ni, and W. Wang, “Application of machine learning algorithms in lane-changing model for intelligent vehicles exiting to off-ramp,” Transp. -Transp. Sci., vol. 17, no. 1, pp. 124–150, Jan. 2021, doi: 10.1080/23249935.2020.1746861

    work page doi:10.1080/23249935.2020.1746861 2021

  25. [26]

    An integrated Empirical Mode Decomposition and Butterworth filter based vehicle trajectory reconstruction method,

    S. Dong, Y. Zhou, T. Chen, S. Li, Q. Gao, and B. Ran, “An integrated Empirical Mode Decomposition and Butterworth filter based vehicle trajectory reconstruction method,” Phys. -Stat. Mech. ITS Appl., vol. 583, Dec. 2021, doi: 10.1016/j.physa.2021.126295

    work page doi:10.1016/j.physa.2021.126295 2021

  26. [27]

    Ubiquitous Traffic Eyes: trajectory dataset focus on multiple traffic states and state transition on urban expressways,

    R. Feng, H. Zhu, N. Sze, S. Wang, and Z. Li, “Ubiquitous Traffic Eyes: trajectory dataset focus on multiple traffic states and state transition on urban expressways,” Transp. Lett.- Int. J. Transp. Res., vol. 18, no. 2, pp. 446–462, Feb. 2026, doi: 10.1080/19427867.2025.2559276

    work page doi:10.1080/19427867.2025.2559276 2026