Pith. sign in

REVIEW 4 major objections 4 minor 40 references

LoRFT: Benchmarking Long-Range Vehicle Trajectory Reconstruction from Fixed Highway Cameras

T0 review · 4 major / 4 minor · reviewed 2026-08-01 · deepseek-v4-flash

Pith's one-line read A map-aware model reconstructs distant vehicle trajectories in fixed highway camera views, cutting final error by 15.4%, on a new public benchmark.

desk verdict LoRFT is a genuinely novel benchmark with a clean protocol; the reported gains are credible but sit close to an unquantified annotation-noise floor. read the letter →

arxiv 2607.19911 v1 pith:JS5ZEAM5 submitted 2026-07-22 cs.CV

classification cs.CV
keywords vehicletrajectoryreconstructionfixedhighwaycamerasbenchmarkroadgeometrysequence-to-sequencesurveillancevideoimputationfar-rangetracking
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper introduces LoRFT, a benchmark for reconstructing the far-range continuation of vehicle trajectories from fixed highway surveillance cameras, and Map-RSTNet, a model that integrates road geometry into a sequence-to-sequence framework. The model represents trajectories in a road-aligned state space—expressing position and scale relative to road boundaries—and refreshes local geometry during decoding. On LoRFT, Map-RSTNet reduces average displacement error by 11.0%, final displacement error by 15.4%, and 5-second RMSE by 10.5% versus the strongest baseline. The authors argue that fixed cameras can extend usable trajectory records when far-range track fragmentation is treated as a reconstruction problem.

What carries the argument

The key mechanism is the road-aligned state space: each trajectory point is expressed as a lateral fraction of local road width, a normalized vertical coordinate, scale terms, velocities, and geometry features (distance to left/right boundaries, heading difference from the road centerline). The decoder autoregressively predicts residual displacements in this normalized space, with a hard step limit and a boundary-loss regularizer. During decoding, 'dynamic geometry refresh' re-queries the scene road prior at each predicted position, keeping the reconstructed path aligned with the actual road shape and preventing error accumulation over long horizons.

What would settle it

Re-annotate a random subset of far-range trajectory frames with multiple annotators, compute per-point standard deviations, and compare them to the reported ADE (12.3 px) and FDE (21.7 px). If the annotation noise floor is a large share of those numbers (e.g., >30%), the benchmark cannot confidently distinguish Map-RSTNet from the best baseline; if it is small, the claims stand.

Watch

Extended reading notes

Core claim

The central claim is that road-geometry-aware reconstruction is a practical way to recover the distant segment of a vehicle trajectory from a reliable near-field tracklet within one fixed-camera view. LoRFT provides 6,601 manually verified trajectories, pairing reliable near-field observations with context-verified far-range references, along with road-geometry annotations, scene-level splits, and evaluation scripts. Map-RSTNet encodes trajectories in a road-aligned state space, uses relative residual decoding, and dynamically refreshes road geometry at every decoding step. On the benchmark, this yields consistent improvements over adapted forecasting and imputation baselines, with the large

Load-bearing premise

The manually verified far-range reference annotations are accurate enough that pixel errors of roughly 12–27 truly reflect model performance; if annotator localization noise is a substantial fraction of these errors, the benchmark's ability to rank methods—and the headline 10–15% gains—would be smaller than they appear.

Editorial extensions

If this is right

  • Existing fixed highway camera infrastructure can provide longer trajectory records for safety analysis and traffic management without additional sensors.
  • Road-aligned state representation reduces perspective-induced ambiguity, so models can better extrapolate motion in far-range image regions.
  • LoRFT enables reproducible evaluation of far-range reconstruction, detection, and tracking in the image plane, with scene-level splits testing generalization to unseen views.
  • The largest error reductions at 4–5 seconds suggest that geometry refresh mainly stabilizes long-horizon reconstruction.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the annotation-noise floor is small relative to reported errors, the same road-aligned approach could be transferred to other single-camera settings (urban intersections, toll plazas) where perspective compression fragments tracks.
  • Because LoRFT provides original videos and identity-consistent boxes, it could serve as a testbed for self-supervised or semi-supervised far-range tracking, using reconstructed trajectories as pseudo-labels.
  • The dynamic geometry refresh idea generalizes: any sequence model operating on road scenes could condition on locally refreshed map features rather than a single static encoding.
  • Metric speed and acceleration analyses using reconstructed trajectories will require additional camera calibration; the dataset's image-space format bounds the claims to pixel-space accuracy.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. LoRFT is introduced as the first open benchmark for long-range vehicle trajectory reconstruction from fixed highway cameras, containing 22 expressway scenes, 366,109 frames, 6,601 manually verified trajectories, 2,694,889 bounding boxes, road-geometry annotations, scene-level splits, and evaluation scripts. The paper also proposes Map-RSTNet, a map-aware residual Seq2Seq LSTM that reconstructs a distant trajectory segment from a reliable near-field tracklet in a road-aligned state space with dynamic geometry refresh. The central claim is that Map-RSTNet reduces ADE, FDE, and RMSE@5s by 11.0%, 15.4%, and 10.5%, respectively, relative to the strongest baseline (PRF) on a held-out test set of 4 scenes.

Significance. The benchmark addresses a real gap: existing highway datasets are either overhead or do not supply manually verified far-range continuations of fragmented tracklets. The evaluation protocol is well designed—scene-level splits prevent leakage, labels are manually verified and independent of the proposed model, and the released code and evaluation scripts support reproducibility. The method is sensible, and the ablation is internally consistent with Table 2. The principal weakness is that the benchmark's resolution is not established: the reported improvements are a few pixels, and no label-quality analysis is provided to show that these differences exceed annotation noise. If that gap is closed, the paper would be a solid contribution.

major comments (4)
  1. [Sec. 3.2, Table 2] The central quantitative claim rests on small pixel differences: Map-RSTNet beats PRF by 1.53 px ADE (12.32 vs 13.85), 3.96 px FDE (21.71 vs 25.67), and 3.23 px RMSE@5s (27.47 vs 30.70). The paper provides no inter-annotator agreement, no localization-noise measurement, and no label-error analysis as a function of distance. Given that the paper itself states that distant vehicles are 'small or blurred in a single frame' (Sec. 2.2) and that identities are verified from motion continuity (Sec. 3.2), the reference labels could plausibly carry several pixels of noise. Without a label-quality study, the benchmark cannot be certified as a measurement instrument capable of resolving the reported differences, and the claimed margins might be within the noise floor. Please add re-annotation of a random subset by multiple annotators (reporting per-point standard deviation, box IoU, or similar), a
  2. [Sec. 5.1, Sec. 5.3] All results are reported from a single training run. The absolute differences are small (a few pixels), and deep sequence models can exhibit nontrivial variance across random seeds. Without multiple runs and error bars, or a paired statistical test across test trajectories, the 11–15% relative improvements are not established as beyond run-to-run variation. Please report mean ± standard deviation over at least three seeds and, ideally, a paired test (e.g., Wilcoxon signed-rank over trajectories) for the headline metrics.
  3. [Sec. 5.3] The evaluation protocol says the baselines are 'adapted' to LoRFT, but the manuscript does not specify the adaptation details: input/output representations, whether map information was provided to each baseline, training schedules, hyperparameter searches, or model capacities. If the baselines were not given comparable tuning, the comparison is not a fair test of Map-RSTNet's advantage. Provide complete configuration details or release the baseline code/configs so the comparison can be reproduced and audited.
  4. [Sec. 4.2, Eqs. (5)-(7)] The 10-dimensional state vector contains redundant geometry terms. From Eqs. (4)–(7), ℓL_t = u_t, ℓR_t = 1 − u_t, and f_lat_t = 2u_t − 1. This redundancy does not invalidate the method, but it should be acknowledged and justified (e.g., as an intentionally overcomplete representation) or the redundant terms should be removed for clarity and reproducibility.
minor comments (4)
  1. [Sec. 3.2] The split between observed and reference segments is defined by the reliability of the automatic tracklet, which in this implementation is produced by YOLOv11+ByteTrack. This couples the benchmark's task definition to a particular detector/tracker. Please discuss the potential impact on the benchmark's generality and consider releasing a tracklet-reliability score to allow alternative splits.
  2. [Fig. 5] The hyperparameter curves appear to be on validation data, but the caption says 'on the LoRFT benchmark.' Please clarify whether these are validation or test curves, and add error bars over seeds or state explicitly that they are single-run validation results.
  3. [Sec. 4.2] The notation ℓL_t and ℓR_t, while defined, may confuse readers because they duplicate u_t and 1−u_t. A brief note that these are included as explicit geometry features despite the algebraic redundancy would preempt confusion.
  4. [Sec. 1, Abstract] The claim of being 'the first open benchmark dedicated to long-range vehicle trajectory reconstruction from fixed highway cameras' is plausible but should be supported with a more explicit discussion of why existing datasets (e.g., UA-DETRAC, CHD, V2X-Seq) cannot be repurposed for this exact protocol, rather than only that they were not designed for it.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: benchmark references, road geometry, and method evaluation are mutually independent by construction.

full rationale

The paper's central claims are not circular. LoRFT's ground-truth reference segments are manually verified image-space annotations derived from the videos (Sec. 3.2), independent of Map-RSTNet's predictions. The road-geometry prior M={B,C,Z} is a separately annotated scene-level input, not derived from the reference labels. Map-RSTNet is trained with the objective in Eq. 12 on observed/reference pairs from training scenes, with all hyperparameters selected on validation, and is evaluated on held-out test scenes (Sec. 5.1). Therefore the reported ADE/FDE/RMSE reductions in Table 2 are not fitted to the test labels by construction. The model's output is recovered via Eq. 15 and compared with the reference via Eq. 16; these are distinct quantities, not the same by definition. There is no load-bearing self-citation: the cited prior datasets and methods are background support, and no uniqueness or ansatz is imported solely from the authors' own prior work. The unquantified far-range annotation noise noted in the skeptic's take is a measurement-quality and benchmark-resolution concern, not a circularity, because it does not make any prediction equal its input or any fitted parameter the renamed target. Thus the derivation chain is self-contained and the circularity score is 0.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central method carries no invented entities: the 'scene map' is an annotation product (boundaries/centerlines), not a postulated mechanism. The ledger's weight sits in four validation-tuned hyperparameters (loss weights, step clamp, velocity scale, horizons) and four domain assumptions about annotation quality, tracker-defined task boundaries, far-range map accuracy, and the adequacy of pixel-space error. The free-parameter load is moderate, mostly disclosed, and chosen on validation rather than test — the main gap is the unquantified noise floor of the manual labels and the map.

free parameters (4)
  • Loss weights (λpos=1200, λvel=10, λend=150, λbdry=40) = 1200 / 10 / 150 / 40
    Selected on the validation set (Sec. 5.1) and fixed for test; only λbdry has a partial sensitivity check (Fig. 5d).
  • Decoder step clamp δmax = 0.04
    Hard-clamp on per-step normalized residuals; tuned on validation (Sec. 4.3, Fig. 5b).
  • Velocity scaling factor αv = 10
    Fixed multiplier in velocity features (Sec. 5.1); chosen by hand/validation.
  • Observation horizon; geometry refresh interval = 60 frames; 1 frame
    Observation length chosen via Fig. 5a; refresh interval chosen via Fig. 5c, both on validation.
assumptions (5)
  • domain assumption Far-range vehicle identity can be resolved by annotators from motion continuity across neighboring frames, and the resulting boxes are accurate enough to serve as reconstruction ground truth.
    Underlies the benchmark's reference labels (Fig. 1(b), Sec. 3.2: 'their identities can often be verified from their motion continuity across neighboring frames'). No inter-annotator agreement or label-noise statistic is reported, leaving the ground-truth noise floor unknown.
  • domain assumption The reliability boundary of the YOLOv11 + ByteTrack pipeline is a sensible, stable definition of where the observed segment ends and the reconstruction target begins.
    Sec. 3.2 defines the observed/reference split by 'the reliability of the automatic tracklet.' Any improvement in detection/tracking shifts this boundary, so task difficulty is coupled to one pipeline generation.
  • domain assumption Image-space road-geometry annotations (boundaries, centerlines, zone direction) are accurate in the distant region where reconstruction is evaluated.
    Map-aligned encoding (Eqs. 5-7), dynamic geometry refresh (Sec. 4.4), and pixel-coordinate recovery (Eq. 15) all inherit map errors; no map-quality metric is reported.
  • domain assumption Pixel-space error is an adequate evaluation metric despite the absence of scene-specific camera calibration.
    Adopted in Sec. 5.2; Sec. 6 concedes metric-speed or physical-dynamics applications need calibration. This bounds the benchmark's use for safety-critical claims.
  • standard math Standard background is sound: LSTM Seq2Seq training, Rauch-Tung-Striebel smoothing, Adam optimization, and the L2/smooth-L1 losses behave as intended.
    Invoked throughout Secs. 4-5 without proof; no exotic mathematics is introduced, so this is routine background.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LoRFT: Benchmarking Long-Range Vehicle Trajectory Reconstruction from Fixed Highway Cameras." pith.science (2026). https://pith.science/paper/JS5ZEAM5

@misc{pith2026260719911,
  author       = {Pith},
  title        = {Pith review of: LoRFT: Benchmarking Long-Range Vehicle Trajectory Reconstruction from Fixed Highway Cameras},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JS5ZEAM5}},
  note         = {Machine review of arXiv:2607.19911}
}
read the original abstract

Long-range vehicle trajectories provide important spatio-temporal evidence for traffic safety analysis, autonomous driving evaluation, and data-driven traffic management, yet continuously recovering them from fixed highway cameras remains difficult. As vehicles recede into distant road regions, perspective compression and scale decay often fragment or prematurely terminate automatic tracklets, even when their continuation remains identifiable from motion consistency across neighboring frames. We formulate this problem as recovering the far-range continuation of a vehicle trajectory from a reliable near-field tracklet. We introduce LoRFT, to our knowledge the first open benchmark dedicated to long-range vehicle trajectory reconstruction from fixed highway cameras. LoRFT comprises 22 expressway surveillance scenes, 366,109 video frames, 6,601 manually verified trajectories, 2,694,889 bounding boxes, road-geometry annotations, scene-level splits, and evaluation scripts. We further propose Map-RSTNet, a map-aware residual sequence-to-sequence model that reconstructs distant trajectories in a road-geometry-aligned state space and dynamically refreshes local road geometry during decoding. On LoRFT, Map-RSTNet reduces ADE, FDE, and 5-second RMSE by 11.0%, 15.4%, and 10.5%, respectively, relative to the strongest baseline. These results demonstrate that road-geometry-aware reconstruction can extend usable trajectory records from existing fixed-camera infrastructure. LoRFT provides a reproducible testbed for long-range vehicle trajectory reconstruction.

Figures

Figures reproduced from arXiv: 2607.19911 by the authors.

Figure 1
Figure 1. Motivation and task definition of LoRFT. (a) UAV sensing provides wide coverage [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Spatial distribution of selected LoRFT recording sites on expressways in Sichuan [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Overall framework of Map-RSTNet. Near-field observed tracklets and scene-level road [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison between Map-RSTNet and PRF on a LoRFT test scene. The [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: Impact of key hyperparameters on the LoRFT benchmark. (a) Observation horizon. [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 1 linked inside Pith

  1. [1]

    Itpnet: Towards instanta- neous trajectory prediction for autonomous driving,

    R. Li, C. Li, Y. Li, H. Li, Y. Chen, Y. Yuan, and G. Wang, “Itpnet: Towards instanta- neous trajectory prediction for autonomous driving,” inProc. 30th ACM SIGKDD Conf. Knowledge Discovery and Data Mining (KDD), 2024, pp. 1643–1654. 14

  2. [2]

    Trajectory-LLM: A language-based data generator for trajectory prediction in autonomous driving,

    K. Yang, Z. Guo, G. Lin, H. Dong, Z. Huang, Y. Wu, D. Zuo, J. Peng, Z. Zhong, X. Wang, Q. Guo, X. Jia, J. Yan, and D. Lin, “Trajectory-LLM: A language-based data generator for trajectory prediction in autonomous driving,” inProc. International Conference on Learning Representations (ICLR), 2025

  3. [3]

    A deep transfer learning approach for real-time traffic conflict prediction with trajectory data,

    Q. Hou, Y. Yang, J. Liang, X. Huo, and J. Leng, “A deep transfer learning approach for real-time traffic conflict prediction with trajectory data,”Accident Analysis & Prevention, vol. 214, 2025, art. no. 107966

  4. [4]

    Traffic state estimation from vehicle trajectories with anisotropic gaussian processes,

    F. Wu, Z. Cheng, H. Chen, Z. Qiu, and L. Sun, “Traffic state estimation from vehicle trajectories with anisotropic gaussian processes,”Transportation Research Part C: Emerging Technologies, vol. 163, Jun. 2024, art. no. 104646

  5. [5]

    Traffic light optimization with low penetration rate vehicle trajectory data,

    X. Wang, Z. Jerome, Z. Wang, C. Zhang, S. Shen, V. V. Kumar, F. Bai, P. Krajewski, D. Deneau, A. Jawad, R. Jones, G. Piotrowicz, and H. X. Liu, “Traffic light optimization with low penetration rate vehicle trajectory data,”Nature Communications, vol. 15, no. 1, 2024, art. no. 1306

  6. [6]

    HDSVT: High-density semantic vehicle trajectory dataset based on a cosmopolitan city bridge,

    D. Wen, Y. Zhu, Z. Wu, Z. Wu, Q. Zheng, S. Wang, Y. Wang, and Z. He, “HDSVT: High-density semantic vehicle trajectory dataset based on a cosmopolitan city bridge,” Scientific Data, vol. 13, 2026, art. no. 5

  7. [7]

    I-24 motion: An instrument for freeway traffic science,

    D. Gloudemans, Y. Wang, J. Ji, G. Zach´ ar, W. Barbour, E. Hall, M. Cebelak, L. Smith, and D. B. Work, “I-24 motion: An instrument for freeway traffic science,”Transportation Research Part C: Emerging Technologies, vol. 155, Oct. 2023, art. no. 104311

  8. [8]

    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,” inProc. IEEE 21st Int. Conf. Intelligent Transportation Systems (ITSC), 2018, pp. 2118–2125

Show all 40 references
  1. [9]

    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,” inProc. IEEE Intelligent Vehicles Symposium (IV), 2022, pp. 958–964

  2. [10]

    Transportation management center video recording and archiving best general practices,

    S. Kuciemba and K. Swindler, “Transportation management center video recording and archiving best general practices,” Federal Highway Administration, Tech. Rep. FHW A- HOP-16-033, Mar. 2016

  3. [11]

    Real-time full-stack traffic scene perception for autonomous driving with roadside cameras,

    Z. Zou, R. Zhang, S. Shen, G. Pandey, P. Chakravarty, A. Parchami, and H. X. Liu, “Real-time full-stack traffic scene perception for autonomous driving with roadside cameras,” inProc. IEEE Int. Conf. Robotics and Automation (ICRA), 2022, pp. 890–896

  4. [12]

    Real-time incremental segmentation and tracking of vehicles at low camera angles using stable features,

    N. K. Kanhere and S. T. Birchfield, “Real-time incremental segmentation and tracking of vehicles at low camera angles using stable features,”IEEE Transactions on Intelligent Transportation Systems, vol. 9, no. 1, pp. 148–160, Mar. 2008

  5. [13]

    Rope3d: The roadside perception dataset for autonomous driving and monocular 3d object detection task,

    X. Ye, M. Shu, H. Li, Y. Shi, Y. Li, G. Wang, X. Tan, and E. Ding, “Rope3d: The roadside perception dataset for autonomous driving and monocular 3d object detection task,” in Proc. IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), 2022, pp. 21 341–21 350

  6. [14]

    Ua- detrac: A new benchmark and protocol for multi-object detection and tracking,

    L. Wen, D. Du, Z. Cai, Z. Lei, M.-C. Chang, H. Qi, J. Lim, M.-H. Yang, and S. Lyu, “Ua- detrac: A new benchmark and protocol for multi-object detection and tracking,”Computer Vision and Image Understanding, vol. 193, Apr. 2020, art. no. 102907. 15

  7. [15]

    A pov-based highway vehicle trajectory dataset and prediction architecture,

    V. Katariya, G. Alinezhad Noghre, A. Danesh Pazho, and H. Tabkhi, “A pov-based highway vehicle trajectory dataset and prediction architecture,”IEEE Transactions on Intelligent Transportation Systems, vol. 25, no. 10, pp. 13 136–13 146, Oct. 2024

  8. [16]

    The next generation simulation program,

    V. Alexiadis, J. Halkias, G. McHale, J. Colyar, and R. Hranac, “The next generation simulation program,”ITE Journal, vol. 74, no. 8, pp. 22–26, 2004

  9. [17]

    V2x-seq: A large-scale sequential dataset for vehicle- infrastructure cooperative perception and forecasting,

    H. Yu, W. Yang, H. Ruan, Z. Yang, Y. Tang, X. Gao, X. Hao, Y. Shi, Y. Pan, N. Sun, J. Song, J. Yuan, P. Luo, and Z. Nie, “V2x-seq: A large-scale sequential dataset for vehicle- infrastructure cooperative perception and forecasting,” inProc. IEEE/CVF Conf. Computer Vision and P...

  10. [18]

    The ind dataset: A drone dataset of naturalistic road user trajectories at german intersections,

    J. Bock, R. Krajewski, T. Moers, S. Runde, L. Vater, and L. Eckstein, “The ind dataset: A drone dataset of naturalistic road user trajectories at german intersections,” inProc. IEEE Intelligent Vehicles Symposium (IV), 2020, pp. 1929–1934

  11. [19]

    The round dataset: A drone dataset of road user trajectories at roundabouts in germany,

    R. Krajewski, T. Moers, J. Bock, L. Vater, and L. Eckstein, “The round dataset: A drone dataset of road user trajectories at roundabouts in germany,” inProc. IEEE 23rd Int. Conf. Intelligent Transportation Systems (ITSC), 2020, pp. 1–6

  12. [20]

    opendd: A large-scale roundabout drone dataset,

    A. Breuer, J.-A. Termoehlen, S. Homoceanu, and T. Fingscheidt, “opendd: A large-scale roundabout drone dataset,” inProc. IEEE 23rd Int. Conf. Intelligent Transportation Systems (ITSC), 2020, pp. 1–6

  13. [21]

    On the new era of urban traffic monitoring with massive drone data: The pneuma large-scale field experiment,

    E. Barmpounakis and N. Geroliminis, “On the new era of urban traffic monitoring with massive drone data: The pneuma large-scale field experiment,”Transportation Research Part C: Emerging Technologies, vol. 111, pp. 50–71, Feb. 2020

  14. [22]

    Vehicle trajectory prediction based on motion model and maneuver recognition,

    A. Houenou, P. Bonnifait, V. Cherfaoui, and W. Yao, “Vehicle trajectory prediction based on motion model and maneuver recognition,” inProc. IEEE/RSJ Int. Conf. Intelligent Robots and Systems (IROS), 2013, pp. 4363–4369

  15. [23]

    Driving intention recognition and behaviour prediction based on a double-layer hidden markov model,

    L. He, C. Zong, and C. Wang, “Driving intention recognition and behaviour prediction based on a double-layer hidden markov model,”Journal of Zhejiang University-SCIENCE C, vol. 13, no. 3, pp. 208–217, Mar. 2012

  16. [24]

    An integrated approach to maneuver-based trajectory prediction and criticality assessment in arbitrary road environments,

    M. Schreier, V. Willert, and J. Adamy, “An integrated approach to maneuver-based trajectory prediction and criticality assessment in arbitrary road environments,”IEEE Transactions on Intelligent Transportation Systems, vol. 17, no. 10, pp. 2751–2766, Oct. 2016

  17. [25]

    Machine learning for autonomous vehicle’s trajectory pre- diction: A comprehensive survey, challenges, and future research directions,

    V. Bharilya and N. Kumar, “Machine learning for autonomous vehicle’s trajectory pre- diction: A comprehensive survey, challenges, and future research directions,”Vehicular Communications, vol. 46, Apr. 2024, art. no. 100733

  18. [26]

    Convolutional social pooling for vehicle trajectory prediction,

    N. Deo and M. M. Trivedi, “Convolutional social pooling for vehicle trajectory prediction,” inProc. IEEE/CVF Conf. Computer Vision and Pattern Recognition Workshops (CVPR W), 2018, pp. 1468–1476

  19. [27]

    Deeptrack: Lightweight deep learning for vehicle trajectory prediction in highways,

    V. Katariya, M. Baharani, N. Morris, O. Shoghli, and H. Tabkhi, “Deeptrack: Lightweight deep learning for vehicle trajectory prediction in highways,”IEEE Transactions on Intelligent Transportation Systems, vol. 23, no. 10, pp. 18 927–18 936, Oct. 2022

  20. [28]

    Grip++: Enhanced graph-based interaction-aware trajectory prediction for autonomous driving,

    X. Li, X. Ying, and M. C. Chuah, “Grip++: Enhanced graph-based interaction-aware trajectory prediction for autonomous driving,” arXiv preprint arXiv:1907.07792, 2020. 16

  21. [29]

    Scale-net: Scalable vehicle trajectory prediction network under random number of interacting vehicles via edge-enhanced graph convolutional neural network,

    H. Jeon, J. Choi, and D. Kum, “Scale-net: Scalable vehicle trajectory prediction network under random number of interacting vehicles via edge-enhanced graph convolutional neural network,” inProc. IEEE/RSJ Int. Conf. Intelligent Robots and Systems (IROS), 2020, pp. 2095–2102

  22. [30]

    Densetnt: End-to-end trajectory prediction from dense goal sets,

    J. Gu, C. Sun, and H. Zhao, “Densetnt: End-to-end trajectory prediction from dense goal sets,” inProc. IEEE/CVF Int. Conf. Computer Vision (ICCV), 2021, pp. 15 283–15 292

  23. [31]

    Learning lane graph representations for motion forecasting,

    M. Liang, B. Yang, R. Hu, Y. Chen, R. Liao, S. Feng, and R. Urtasun, “Learning lane graph representations for motion forecasting,” inComputer Vision – ECCV 2020, ser. Lecture Notes in Computer Science. Cham, Switzerland: Springer, 2020, vol. 12347, pp. 541–556

  24. [32]

    Vectornet: Encoding hd maps and agent dynamics from vectorized representation,

    J. Gao, C. Sun, H. Zhao, Y. Shen, D. Anguelov, C. Li, and C. Schmid, “Vectornet: Encoding hd maps and agent dynamics from vectorized representation,” inProc. IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), 2020, pp. 11 525–11 533

  25. [33]

    Mixnet: Physics constrained deep neural motion prediction for autonomous racing,

    P. Karle, F. Toeroek, M. Geisslinger, and M. Lienkamp, “Mixnet: Physics constrained deep neural motion prediction for autonomous racing,”IEEE Access, vol. 11, pp. 85 914–85 926, 2023

  26. [34]

    Goal- based neural physics vehicle trajectory prediction model,

    R. Gan, H. Shi, P. Li, K. Wu, B. An, J. You, L. Li, J. Ma, C. Ma, and B. Ran, “Goal- based neural physics vehicle trajectory prediction model,”Transportation Research Part C: Emerging Technologies, vol. 179, Oct. 2025, art. no. 105283

  27. [35]

    Uncovering the missing pattern: Unified framework towards trajectory imputation and prediction,

    Y. Xu, A. Bazarjani, H. Chi, C. Choi, and Y. Fu, “Uncovering the missing pattern: Unified framework towards trajectory imputation and prediction,” inProc. IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), 2023, pp. 9632–9643

  28. [36]

    Automatic vehicle trajectory data reconstruction at scale,

    Y. Wang, D. Gloudemans, J. Ji, Z. N. Teoh, L. Liu, G. Zach´ ar, W. Barbour, and D. Work, “Automatic vehicle trajectory data reconstruction at scale,”Transportation Research Part C: Emerging Technologies, vol. 160, Mar. 2024, art. no. 104520

  29. [37]

    Trajrecovery: An efficient vehicle trajectory recovery framework based on urban-scale traffic camera records,

    D. Wu, Z. Fang, Q. Sun, L. Chen, H. Hu, F. Wang, and Y. Gao, “Trajrecovery: An efficient vehicle trajectory recovery framework based on urban-scale traffic camera records,” in Proc. 30th ACM SIGKDD Conf. Knowledge Discovery and Data Mining (KDD), 2024, pp. 5979–5990

  30. [38]

    A data-driven and kinematics- aware approach to reconstruct full-sample vehicle trajectories from low-quality roadside perception data,

    Z. Zeng, C. Zhao, S. Wang, Y. Ji, C. Wang, and Y. Du, “A data-driven and kinematics- aware approach to reconstruct full-sample vehicle trajectories from low-quality roadside perception data,”IEEE Transactions on Intelligent Transportation Systems, vol. 26, no. 10, pp. 15 041–1...

  31. [39]

    Bytetrack: Multi-object tracking by associating every detection box,

    Y. Zhang, P. Sun, Y. Jiang, D. Yu, Z. Yuan, P. Luo, W. Liu, and X. Wang, “Bytetrack: Multi-object tracking by associating every detection box,” inComputer Vision – ECCV 2022, ser. Lecture Notes in Computer Science. Cham, Switzerland: Springer, 2022, vol. 13682, pp. 1–21

  32. [40]

    Recover to Predict: Progressive retrospective learning for variable-length trajectory prediction,

    H. Zhou, L. Qi, X. Li, J. Zhang, Y. Liu, X. Yang, M. Fan, and F. Luo, “Recover to Predict: Progressive retrospective learning for variable-length trajectory prediction,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recogn. (CVPR), Jun. 2026, pp. 17 863– 17 873. 17

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.