From Observation to Prediction: LSTM for Vehicle Lane Change Forecasting on Highway On/Off-Ramps

Catherine M. Elias; Mohamed Abouras

arxiv: 2601.14848 · v1 · pith:7B4YNDBWnew · submitted 2026-01-21 · 💻 cs.LG · cs.AI· cs.NE· cs.RO

From Observation to Prediction: LSTM for Vehicle Lane Change Forecasting on Highway On/Off-Ramps

Mohamed Abouras , Catherine M. Elias This is my paper

Pith reviewed 2026-05-21 15:57 UTC · model grok-4.3

classification 💻 cs.LG cs.AIcs.NEcs.RO

keywords LSTMlane change predictionhighway rampsExiD datasetvehicle trajectory forecastingtraffic behavior modelingmachine learning for safety

0 comments

The pith

LSTM networks trained on drone footage predict lane changes on highway on and off ramps up to four seconds ahead.

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

The paper examines whether multi-layered LSTM models can forecast when vehicles will change lanes in the complex traffic patterns near highway on and off ramps. These ramp sections create more variation than straight highway stretches, so reliable short-term predictions could help reduce collisions and improve flow. The authors train and test the models on the ExiD drone dataset, comparing results for the ramp areas of interest against general highway cases. They evaluate several prediction horizons and model setups, finding that accuracy remains usable out to the longest horizon examined. A reader might care because such forecasts could feed into driver assistance systems or traffic management tools that react before a maneuver begins.

Core claim

The paper shows that a multi-layered LSTM architecture trained on the ExiD drone dataset can forecast lane change maneuvers in areas of interest such as highway on and off-ramps. The study compares this to straight highway sections and finds promising results with prediction accuracies reaching approximately 76 percent for the ramp areas and 94 percent for general highway scenarios at the maximum prediction horizon of four seconds.

What carries the argument

Multi-layered LSTM architecture that processes vehicle trajectory sequences to output lane-change probabilities at chosen future horizons.

If this is right

Predictions at four-second horizons can lower uncertainty in ramp interactions compared with shorter or longer windows.
The same LSTM setup produces higher accuracy on straight highway sections than on ramps, confirming the need to treat the two environments separately.
Testing multiple horizons and workflow variants shows that accuracy declines gradually rather than collapsing at longer times.
Drone-collected position data alone suffices to train models that distinguish ramp behavior from ordinary highway driving.

Where Pith is reading between the lines

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

Such models could be combined with real-time camera feeds on vehicles to give earlier warnings to nearby drivers.
Extending the input features to include nearby vehicle speeds or road curvature might raise ramp accuracy closer to the highway numbers.
The four-second horizon opens a window for automated systems to adjust speed or change lanes before the predicted maneuver occurs.

Load-bearing premise

The ExiD drone dataset contains representative samples of real-world lane-change behavior on ramps and the LSTM can generalize to unseen traffic conditions without extra domain-specific rules.

What would settle it

A test on fresh ramp footage not used in training that shows accuracy falling below 60 percent at the four-second horizon would indicate the reported performance does not hold.

Figures

Figures reproduced from arXiv: 2601.14848 by Catherine M. Elias, Mohamed Abouras.

**Figure 1.** Figure 1: Area of Interest existing literature to inform the selection of different prediction horizons. 4) It investigates feature bias across different maneuver classes to understand which inputs most influence model predictions. In section II, this paper will begin by discussing the methodology of this study. Then, in section III, the results of the experiments conducted will be reviewed. Section IV concludes the… view at source ↗

**Figure 4.** Figure 4: Survey Outcome III. RESULTS This section explores the results of the experiments. Firstly, the results of the prediction horizon experiment are recorded. Then, the different stacked LSTM models are evaluated. Finally, the feature bias observation is reviewed. A. Prediction Horizon As the experienced passengers comment on the maneuver, they consider non-empty environments as well in their answers, as per t… view at source ↗

**Figure 3.** Figure 3: Multi-L Models As the number of learning features for the model is substantial, a stack of multiple LSTMs is used; additionally, as the dataset is relatively huge, overfitting must be avoided. Using multiple LSTMs allows the model to extract hierarchical features and increase long-term memory dependency. It also increases the model’s capacity to learn from the given data. Cross-Entropy is used as the loss… view at source ↗

**Figure 5.** Figure 5: Vehicle path example vs Prediction over time. [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗

read the original abstract

On and off-ramps are understudied road sections even though they introduce a higher level of variation in highway interactions. Predicting vehicles' behavior in these areas can decrease the impact of uncertainty and increase road safety. In this paper, the difference between this Area of Interest (AoI) and a straight highway section is studied. Multi-layered LSTM architecture to train the AoI model with ExiD drone dataset is utilized. In the process, different prediction horizons and different models' workflow are tested. The results show great promise on horizons up to 4 seconds with prediction accuracy starting from about 76% for the AoI and 94% for the general highway scenarios on the maximum horizon.

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.

Desk Editor's Note private letter to a colleague

This applies a standard LSTM to ramp lane changes on ExiD data and reports 76-94% accuracies at 4s, but without baselines or imbalance checks those numbers are hard to read as evidence of skill.

read the letter

The punchline for this paper is that it takes an established LSTM model for vehicle trajectory prediction and tests it on the less-studied on and off-ramp sections of highways using the ExiD drone dataset. It finds decent accuracy numbers up to 4 seconds ahead, with 76 percent for the ramp areas and 94 percent for general highway, but the lack of basic controls makes those numbers hard to trust as evidence of real predictive power. What the paper does reasonably well is to draw attention to the higher variation in ramp interactions and to compare performance between the area of interest and straight sections. Testing different prediction horizons and model workflows shows some thought about practical use cases for autonomous driving safety. The choice of drone data is appropriate because it provides detailed trajectory information without the noise of onboard sensors. The soft spots are in the evaluation. The central accuracy claims sit without any baseline comparisons or discussion of class imbalance. Lane change events are typically rare in driving data, often less than 30 percent of sequences. A trivial model that always predicts no lane change would likely achieve 70 to 80 percent accuracy on the same data. Without showing that the LSTM beats such a baseline or a constant velocity predictor, it's impossible to say the model has learned meaningful dynamics. The manuscript also omits details on how the data was split for training and testing, any error bars on the accuracies, or statistical tests to confirm the differences are significant. This is a common issue in applied ML papers but it undercuts the great promise conclusion here. The approach is standard supervised learning on external data, so no circularity problems, and the math is the usual LSTM setup. The contribution is incremental rather than novel, building on existing literature for highway prediction but narrowing to ramps. This paper is aimed at engineers and researchers working on autonomous vehicle behavior prediction in complex road geometries. Someone looking for ideas on applying LSTMs to ramp scenarios might find the horizon tests and dataset useful as a starting point. I would send it to peer review. The topic has relevance for safety applications, and referees could help strengthen the experimental design with the needed baselines and details. It's not ready as is, but it has enough substance to warrant feedback rather than a desk reject.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces a multi-layered LSTM architecture trained on the ExiD drone dataset to predict lane change maneuvers for vehicles on highway on/off-ramps (referred to as Area of Interest or AoI) versus straight highway sections. It evaluates the model across different prediction horizons and reports accuracies of approximately 76% for AoI and 94% for general highway scenarios at the maximum 4-second horizon, claiming great promise for these forecasts.

Significance. If the performance claims are substantiated through proper baselines and statistical validation, this work could contribute to safer autonomous vehicle systems by addressing understudied ramp interactions. The explicit comparison of AoI versus straight-highway scenarios is a useful framing. However, without evidence that the LSTM outperforms trivial predictors, the practical significance for forecasting remains difficult to assess.

major comments (2)

[Abstract] Abstract: The abstract states accuracy numbers for different horizons but supplies no information on training/validation splits, baseline comparisons, error bars, or statistical tests, leaving the central performance claim difficult to evaluate.
[Results] Results section: The reported accuracies (starting from 76% for AoI and 94% for highway at the 4 s horizon) are presented without class ratios, majority-class baseline, or comparison to simpler models such as constant-velocity or logistic regression on hand-crafted features. Given that lane-change events are typically rare, this omission prevents determining whether the LSTM delivers genuine forecasting skill.

minor comments (2)

The phrase 'different models' workflow' is used without a clear definition or diagram; a flowchart or pseudocode would improve reproducibility.
The manuscript should explicitly state the LSTM hyperparameters (layers, hidden units, dropout) and the exact input feature representation from the drone trajectories.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive report. The comments highlight important aspects for strengthening the evaluation of our LSTM-based lane change prediction approach. We address each major comment below and will incorporate revisions to improve clarity and substantiation of the results.

read point-by-point responses

Referee: [Abstract] Abstract: The abstract states accuracy numbers for different horizons but supplies no information on training/validation splits, baseline comparisons, error bars, or statistical tests, leaving the central performance claim difficult to evaluate.

Authors: We agree that additional context in the abstract would aid evaluation. In the revised manuscript, we will expand the abstract to note the 70/15/15 training/validation/test split on the ExiD dataset, report mean accuracies with standard deviations across multiple runs, and reference that statistical significance was evaluated using paired t-tests against baselines. This addresses the concern while preserving the abstract's focus on the AoI versus straight-highway comparison. revision: yes
Referee: [Results] Results section: The reported accuracies (starting from 76% for AoI and 94% for highway at the 4 s horizon) are presented without class ratios, majority-class baseline, or comparison to simpler models such as constant-velocity or logistic regression on hand-crafted features. Given that lane-change events are typically rare, this omission prevents determining whether the LSTM delivers genuine forecasting skill.

Authors: We acknowledge that explicit baselines are necessary to demonstrate forecasting skill beyond trivial predictors, especially given class imbalance. The manuscript emphasizes performance differences between AoI and straight sections but does not prominently feature these elements. We will revise the results section to report class ratios (lane changes comprise roughly 12-18% of samples depending on the area), include a majority-class baseline, and add comparisons to a constant-velocity model and logistic regression using features such as speed, acceleration, and relative position. Updated tables will show that the multi-layered LSTM outperforms these baselines, particularly at longer horizons. revision: yes

Circularity Check

0 steps flagged

No circularity: standard supervised LSTM training on external dataset

full rationale

The paper applies a conventional supervised learning pipeline: trajectories from the external ExiD drone dataset are used to train a multi-layer LSTM that outputs lane-change predictions at future horizons. Reported accuracies (76% AoI, 94% highway at 4 s) are empirical test-set metrics, not quantities that reduce by the paper's own equations to parameters fitted from the same target labels. No self-definitional steps, fitted-input-as-prediction, or load-bearing self-citations appear; the derivation chain remains independent of the evaluation numbers themselves.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the representativeness of the ExiD drone trajectories and the capacity of a standard recurrent network to capture temporal dependencies in lane-change behavior.

free parameters (1)

LSTM hyperparameters
Number of layers, hidden units, learning rate, and sequence length are selected and optimized during training on the dataset.

axioms (1)

domain assumption The ExiD drone dataset accurately represents real-world vehicle positions and lane changes on both straight highways and on/off-ramps
Training and evaluation are performed directly on this dataset without reported cross-validation against independent sources.

pith-pipeline@v0.9.0 · 5653 in / 1247 out tokens · 65098 ms · 2026-05-21T15:57:34.909552+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Multi-layered LSTM architecture... prediction accuracy starting from about 76% for the AoI and 94% for the general highway scenarios on the maximum horizon.
IndisputableMonolith/Foundation/ArithmeticFromLogic.lean LogicNat induction and 8-tick period unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

observation period of five frames... different prediction horizons

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

16 extracted references · 16 canonical work pages

[1]

More than a million people die from road injuries every year,

H. Ritchie, “More than a million people die from road injuries every year,” https://ourworldindata.org/data-insights/ more-than-a-million-people-die-from-road-injuries-every-year, 2024, accessed: 2025-07

work page 2024
[2]

Road traffic injuries,

World Health Organization, “Road traffic injuries,” https://www.who. int/news-room/fact-sheets/detail/road-traffic-injuries, Dec. 2023, ac- cessed: 2025-07

work page 2023
[3]

Machine learning for autonomous vehi- cle’s trajectory prediction: A comprehensive survey, challenges, and future research directions,

V . Bharilya and N. Kumar, “Machine learning for autonomous vehi- cle’s trajectory prediction: A comprehensive survey, challenges, and future research directions,”Vehicular Communications, vol. 46, p. 100733, 2024

work page 2024
[4]

An enhanced vehicle trajectory prediction model leveraging lstm and social-attention mechanisms,

S. Qiao, F. Gao, J. Wu, and R. Zhao, “An enhanced vehicle trajectory prediction model leveraging lstm and social-attention mechanisms,” IEEE Access, vol. 12, pp. 1718–1726, 2024

work page 2024
[5]

Vehicle lane change prediction based on knowledge graph embeddings and bayesian inference,

M. Manzour, A. Ballardini, R. Izquierdo, and M. Sotelo, “Vehicle lane change prediction based on knowledge graph embeddings and bayesian inference,” in2024 IEEE Intelligent Vehicles Symposium (IV). IEEE, 2024, pp. 1893–1900

work page 2024
[6]

Vehicle trajectory prediction in top-view im- age sequences based on deep learning method,

Z. S. Nejad, H. Heravi, A. R. Jounghani, A. Shahrezaie, and A. Ebrahimi, “Vehicle trajectory prediction in top-view im- age sequences based on deep learning method,”arXiv preprint arXiv:2102.01749, 2021

work page arXiv 2021
[7]

Trajectory-prediction with vision: A survey,

A. Singh, “Trajectory-prediction with vision: A survey,” inProceed- ings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 3318–3323

work page 2023
[8]

A review of deep learning-based vehicle motion prediction for autonomous driving,

R. Huang, G. Zhuo, L. Xiong, S. Lu, and W. Tian, “A review of deep learning-based vehicle motion prediction for autonomous driving,”Sustainability, vol. 15, no. 20, 2023. [Online]. Available: https://www.mdpi.com/2071-1050/15/20/14716

work page 2023
[9]

A survey on deep-learning approaches for vehicle trajectory prediction in autonomous driving,

J. Liu, X. Mao, Y . Fang, D. Zhu, and M. Q. Meng, “A survey on deep-learning approaches for vehicle trajectory prediction in autonomous driving,”CoRR, vol. abs/2110.10436, 2021. [Online]. Available: https://arxiv.org/abs/2110.10436

work page arXiv 2021
[10]

Artificial intelligence techniques for driving safety and vehicle crash prediction,

Z. Halim, R. Kalsoom, S. Bashir, and G. Abbas, “Artificial intelligence techniques for driving safety and vehicle crash prediction,”Artificial Intelligence Review, vol. 46, 10 2016

work page 2016
[11]

INTERACTION dataset: An international, adversarial and cooperative motion dataset in interactive driving scenarios with semantic maps,

W. Zhan, L. Sun, D. Wang, H. Shi, A. Clausse, M. Naumann, J. K ¨ummerle, H. K ¨onigshof, C. Stiller, A. de La Fortelle, and M. Tomizuka, “INTERACTION Dataset: An INTERnational, Ad- versarial and Cooperative moTION Dataset in Interactive Driving Scenarios with Semantic Maps,”arXiv:1910.03088 [cs, eess], Sep. 2019

work page arXiv 1910
[12]

A human factors approach to validating driver models for interaction-aware automated vehicles,

O. Siebinga, A. Zgonnikov, and D. Abbink, “A human factors approach to validating driver models for interaction-aware automated vehicles,” ACM Transactions on Human-Robot Interaction (THRI), vol. 11, no. 4, pp. 1–21, 2022

work page 2022
[13]

Lstm-based preceding vehicle behaviour prediction during aggressive lane change for acc application,

R. Singh, S. Mozaffari, M. Rezaei, and S. Alirezaee, “Lstm-based preceding vehicle behaviour prediction during aggressive lane change for acc application,” in2023 International Symposium on Signals, Circuits and Systems (ISSCS), 2023, pp. 1–4

work page 2023
[14]

A novel model for driver lane change prediction in cooperative adaptive cruise control systems,

A. N. Qasemabadi, S. Mozaffari, M. Rezaei, M. Ahmadi, and S. Alirezaee, “A novel model for driver lane change prediction in cooperative adaptive cruise control systems,” in2023 International Symposium on Signals, Circuits and Systems (ISSCS), 2023, pp. 1–4

work page 2023
[15]

The exid dataset: A real-world trajectory dataset of highly interactive highway scenarios in germany,

T. Moers, L. Vater, R. Krajewski, J. Bock, A. Zlocki, and L. Eckstein, “The exid dataset: A real-world trajectory dataset of highly interactive highway scenarios in germany,” in2022 IEEE Intelligent Vehicles Symposium (IV), 2022, pp. 958–964

work page 2022
[16]

The highd dataset: A drone dataset of naturalistic vehicle trajectories on german highways for validation of highly automated driving systems,

R. Krajewski, J. Bock, L. Kloeker, and L. Eckstein, “The highd dataset: A drone dataset of naturalistic vehicle trajectories on german highways for validation of highly automated driving systems,” in2018 21st International Conference on Intelligent Transportation Systems (ITSC), 2018, pp. 2118–2125

work page 2018

[1] [1]

More than a million people die from road injuries every year,

H. Ritchie, “More than a million people die from road injuries every year,” https://ourworldindata.org/data-insights/ more-than-a-million-people-die-from-road-injuries-every-year, 2024, accessed: 2025-07

work page 2024

[2] [2]

Road traffic injuries,

World Health Organization, “Road traffic injuries,” https://www.who. int/news-room/fact-sheets/detail/road-traffic-injuries, Dec. 2023, ac- cessed: 2025-07

work page 2023

[3] [3]

Machine learning for autonomous vehi- cle’s trajectory prediction: A comprehensive survey, challenges, and future research directions,

V . Bharilya and N. Kumar, “Machine learning for autonomous vehi- cle’s trajectory prediction: A comprehensive survey, challenges, and future research directions,”Vehicular Communications, vol. 46, p. 100733, 2024

work page 2024

[4] [4]

An enhanced vehicle trajectory prediction model leveraging lstm and social-attention mechanisms,

S. Qiao, F. Gao, J. Wu, and R. Zhao, “An enhanced vehicle trajectory prediction model leveraging lstm and social-attention mechanisms,” IEEE Access, vol. 12, pp. 1718–1726, 2024

work page 2024

[5] [5]

Vehicle lane change prediction based on knowledge graph embeddings and bayesian inference,

M. Manzour, A. Ballardini, R. Izquierdo, and M. Sotelo, “Vehicle lane change prediction based on knowledge graph embeddings and bayesian inference,” in2024 IEEE Intelligent Vehicles Symposium (IV). IEEE, 2024, pp. 1893–1900

work page 2024

[6] [6]

Vehicle trajectory prediction in top-view im- age sequences based on deep learning method,

Z. S. Nejad, H. Heravi, A. R. Jounghani, A. Shahrezaie, and A. Ebrahimi, “Vehicle trajectory prediction in top-view im- age sequences based on deep learning method,”arXiv preprint arXiv:2102.01749, 2021

work page arXiv 2021

[7] [7]

Trajectory-prediction with vision: A survey,

A. Singh, “Trajectory-prediction with vision: A survey,” inProceed- ings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 3318–3323

work page 2023

[8] [8]

A review of deep learning-based vehicle motion prediction for autonomous driving,

R. Huang, G. Zhuo, L. Xiong, S. Lu, and W. Tian, “A review of deep learning-based vehicle motion prediction for autonomous driving,”Sustainability, vol. 15, no. 20, 2023. [Online]. Available: https://www.mdpi.com/2071-1050/15/20/14716

work page 2023

[9] [9]

A survey on deep-learning approaches for vehicle trajectory prediction in autonomous driving,

J. Liu, X. Mao, Y . Fang, D. Zhu, and M. Q. Meng, “A survey on deep-learning approaches for vehicle trajectory prediction in autonomous driving,”CoRR, vol. abs/2110.10436, 2021. [Online]. Available: https://arxiv.org/abs/2110.10436

work page arXiv 2021

[10] [10]

Artificial intelligence techniques for driving safety and vehicle crash prediction,

Z. Halim, R. Kalsoom, S. Bashir, and G. Abbas, “Artificial intelligence techniques for driving safety and vehicle crash prediction,”Artificial Intelligence Review, vol. 46, 10 2016

work page 2016

[11] [11]

INTERACTION dataset: An international, adversarial and cooperative motion dataset in interactive driving scenarios with semantic maps,

W. Zhan, L. Sun, D. Wang, H. Shi, A. Clausse, M. Naumann, J. K ¨ummerle, H. K ¨onigshof, C. Stiller, A. de La Fortelle, and M. Tomizuka, “INTERACTION Dataset: An INTERnational, Ad- versarial and Cooperative moTION Dataset in Interactive Driving Scenarios with Semantic Maps,”arXiv:1910.03088 [cs, eess], Sep. 2019

work page arXiv 1910

[12] [12]

A human factors approach to validating driver models for interaction-aware automated vehicles,

O. Siebinga, A. Zgonnikov, and D. Abbink, “A human factors approach to validating driver models for interaction-aware automated vehicles,” ACM Transactions on Human-Robot Interaction (THRI), vol. 11, no. 4, pp. 1–21, 2022

work page 2022

[13] [13]

Lstm-based preceding vehicle behaviour prediction during aggressive lane change for acc application,

R. Singh, S. Mozaffari, M. Rezaei, and S. Alirezaee, “Lstm-based preceding vehicle behaviour prediction during aggressive lane change for acc application,” in2023 International Symposium on Signals, Circuits and Systems (ISSCS), 2023, pp. 1–4

work page 2023

[14] [14]

A novel model for driver lane change prediction in cooperative adaptive cruise control systems,

A. N. Qasemabadi, S. Mozaffari, M. Rezaei, M. Ahmadi, and S. Alirezaee, “A novel model for driver lane change prediction in cooperative adaptive cruise control systems,” in2023 International Symposium on Signals, Circuits and Systems (ISSCS), 2023, pp. 1–4

work page 2023

[15] [15]

The exid dataset: A real-world trajectory dataset of highly interactive highway scenarios in germany,

T. Moers, L. Vater, R. Krajewski, J. Bock, A. Zlocki, and L. Eckstein, “The exid dataset: A real-world trajectory dataset of highly interactive highway scenarios in germany,” in2022 IEEE Intelligent Vehicles Symposium (IV), 2022, pp. 958–964

work page 2022

[16] [16]

The highd dataset: A drone dataset of naturalistic vehicle trajectories on german highways for validation of highly automated driving systems,

R. Krajewski, J. Bock, L. Kloeker, and L. Eckstein, “The highd dataset: A drone dataset of naturalistic vehicle trajectories on german highways for validation of highly automated driving systems,” in2018 21st International Conference on Intelligent Transportation Systems (ITSC), 2018, pp. 2118–2125

work page 2018