Pith. sign in

REVIEW 4 major objections 6 minor 75 references

Trajectory Map-Matching in Urban Road Networks Based on RSS Measurements

T0 review · 4 major / 6 minor · reviewed 2026-08-09 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read RSS measurements alone can reconstruct urban vehicle trajectories to about 12–15 m accuracy.

desk verdict The 12 m RSS-only trajectory claim is not supported as written: Eq. (2)'s variance is undefined at the paper's own parameters, though the core idea is new and worth refereeing. read the letter →

arxiv 2502.01280 v2 pith:R4JVVVGF submitted 2025-02-03 eess.SY cs.SY

Zheng Xing , Weibing Zhao This is my paper · ORCID
classification eess.SYcs.SY
keywords RSS-basedlocalizationvehicletrajectoryreconstructionhiddenMarkovmodelmapmatchingroadnetworkreceivedsignalstrengthalternatingoptimization5GNR
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

This paper tries to establish that received signal strength (RSS) readings from cellular base stations, which any phone can record, contain enough information to reconstruct a vehicle's path through an urban road network without GPS, cameras, or specialized time-difference-of-arrival receivers. The key move is to treat trajectory recovery as a hidden Markov model whose states are candidate points on a road graph and whose observations are RSS values from surrounding base stations, governed by a log-distance path-loss law with unknown per-station parameters. An alternating optimization estimates those parameters in closed form, then decodes the most probable trajectory through the graph; a maximum-speed-constrained rough path supplies the initial guess and steers the optimization away from poor local solutions. On real 5G NR measurements from Chengdu and Shenzhen, the paper reports queuing-location errors of 12.6 m and 14.7 m and trajectory-matching errors below 0.4%, far below the map-matching baselines, and robustness to randomly missing RSS data up to about 30%. If true, this would make low-cost, large-scale traffic-flow reconstruction from ordinary cellular signaling a practical option.

What carries the argument

The central object is the hidden Markov model over a road graph $G=(V,E)$: graph nodes are candidate positions spaced $\gamma$ meters apart on roads, and edges connect nodes that a vehicle traveling at no more than speed $v_{\max}$ could cover in one $\delta$-second time slot. The observation probability for an RSS vector is built from a log-distance path-loss model with unknown per-base-station parameters $(\alpha_q,\beta_q,\sigma_q)$, and the transition probability is a Gaussian speed model. The HRE algorithm alternates between a closed-form least-squares fit of the propagation parameters (given the trajectory) and a two-stage forward-backward decode of the trajectory (given the parameters), initialized by MSR, a convex speed-constrained smoothing of RSS-weighted base-station positions. This machinery lets the method enforce both signal-propagation rules and vehicle mobility constraints while searching over the road network.

What would settle it

Take the same Chengdu and Shenzhen RSS traces and rerun HRE with each base-station coordinate perturbed by a zero-mean error with standard deviation of 10–30 meters; if the reported ~12–15 m QLE degrades to tens of meters, then exact tower coordinates, not the RSS alone, carry the claimed accuracy.

Watch

Extended reading notes

Core claim

The central claim is that RSS-only trajectory reconstruction does not need dense, clean location estimates: the noisy and sporadic RSS sequence carries a latent spatiotemporal structure that can be extracted by coupling a physical signal-propagation model with a road-network constraint. The paper's HMM-based RSS embedding (HRE) formalizes this by searching for the node sequence on a road graph that maximizes the joint probability of the RSS observations and the vehicle's motion. The propagation part models each base station's RSS as $\beta_q + \alpha_q \log_{10}\|o_q - x_t\|_2$ plus Gaussian shadowing, with per-station parameters learned by least squares once a candidate trajectory is fixed; the mobility part assigns transition probabilities from a Gaussian speed model whose parameters are either given or, in the adaptive variant HREA, estimated from grouped RSS variation rates. Because the joint problem is non-convex, the method first builds a rough trajectory by solving a convex maximum-speed-constrained smoothing problem, then alternates between closed-form propagation-parameter updates and globally optimal HMM decoding on a coarse-then-fine road graph. On two real 5G NR datasets the paper reports QLE of 12.6 m and 14.7 m (10.1 m and 12.4 m for HREA) and TME of 0.34% and 0.37% (0.28% and 0.25% for HREA), outperforming all compared baselines.

Load-bearing premise

The method assumes the position of every cellular base station is known accurately in advance, but the paper does not say where these coordinates came from or how sensitive the 12–15 m results are to errors in them.

Editorial extensions

If this is right

  • A standard smartphone can serve as the traffic sensor: no specialized decoding receiver or GPS fix is required to reconstruct trajectories at 10–15 m accuracy.
  • Because HRE does not first localize and then match, it avoids the double error the paper attributes to WCL-plus-map-matching pipelines, which is why its reported errors are a factor of 3–5 smaller than the best map-matching baselines.
  • The method tolerates randomly missing RSS: performance stays stable up to roughly 30% missing data and degrades sharply only when fewer than three base-station signals remain per time slot.
  • The adaptive-mobility variant HREA narrows the accuracy gap across road types and traffic densities, making it more suitable for mixed urban driving conditions than the fixed-mobility HRE.
  • The algorithm's runtime grows roughly linearly with trajectory length, so segmenting a long urban route into parallel 1,000-point chunks keeps the method computationally scalable.

Reading between the lines

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

  • The reported accuracy likely depends on precise knowledge of base-station coordinates; the paper neither states how these were obtained nor tests sensitivity to tower-position error, so the practical deployment claim should be read as conditional on that undisclosed data quality.
  • Because the observation model is generic log-distance attenuation, the same alternating HMM embedding should transfer to other ubiquitous emitters such as WiFi access points or IoT gateways, provided their positions and a road graph are available; this is a testable extension the paper does not run.
  • The robustness experiments use random missing masks, whereas real RSS sparsity is often bursty (tunnels, urban canyons, handover gaps), so a fair stress test would evaluate HRE under consecutive missing intervals.
  • The paper's own ablation with known propagation parameters gives QLE of 8.4 m, suggesting that roughly 4 m of the remaining error comes from learning the propagation model rather than from decoding, so better parameter estimation is the most direct path to further gains.
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 / 6 minor

Summary. The paper proposes an RSS-only vehicle trajectory reconstruction (VTR) method for urban road networks. It models RSS observations through a per-BS log-distance path-loss model and vehicle motion through a Gaussian speed model, then casts trajectory recovery as an HMM decoding problem on a road graph. The authors introduce an alternating optimization procedure (HRE) that jointly estimates propagation parameters and the trajectory, a maximum-speed-constrained rough initialization (MSR), and an adaptive variant (HREA) that also estimates mobility parameters per speed group. The method is evaluated on two real 5G NR datasets collected in Chengdu and Shenzhen, with claimed QLE values of 12.6 m and 14.7 m and TME values below 0.4%, substantially outperforming map-matching baselines.

Significance. If the results are reproducible, this is a practically valuable contribution: RSS is available on any standard mobile device, so RSS-only VTR could provide low-cost trajectory reconstruction without specialized TDoA receivers or GNSS. The real-data evaluation over more than 300 km of urban driving is a genuine strength, as is the explicit ablation with ground-truth propagation parameters (Table VI), which gives an upper bound and shows that the alternating procedure is EM-like rather than definitionally circular. The robustness experiments with missing data and noise are also useful. However, several internal inconsistencies in the model equations and algorithm listings currently prevent the empirical claims from being accepted as stated; these need to be resolved before the contribution can be assessed fairly.

major comments (4)
  1. [Section II-C, Eq. (2)] The transition variance is undefined for the reported experimental settings. With vmax=22.2 m/s, vavr=10.5 m/s (Dataset I), and η=0.05, the argument of W_{-1} is -2π(0.05)^2(11.7)^2 ≈ -2.15, which lies outside the real domain [-1/e, 0) ≈ [-0.3679, 0) of the Lambert W branch W_{-1}. Hence no real σ̂_v^2 exists and p(x_t|x_{t-1}) in Eq. (2) cannot be evaluated. Figure 7(a) sweeps η up to 0.08, a range that is invalid above approximately η ≈ 0.021. Since Eq. (2) enters the objective (4) and Algorithm 2, the QLE/TME numbers in Table III cannot be traced to a well-defined transition model. Please correct the model or clarify the alternate formula actually used in the experiments.
  2. [Section IV-C, Algorithm 2] Algorithm 2 selects k̂ = argmin_k l(k) and the final path as argmin_j L(T,j), while the text in Section IV-C describes maximizing the highest score and the highest-probability state. Since the initialized L(1,j) and the increments l(k) are log-probabilities, which are nonpositive, argmin chooses the most unlikely continuation rather than the most likely one. This is a sign inconsistency in the core Viterbi-style decoding step used by both HRE and HREA. Please change argmin to argmax or, alternatively, define L as a negative log-likelihood and propagate that sign convention consistently through all steps.
  3. [Section IV-F, Eq. (16)] The interior-point barrier in Eq. (16) uses −log(−||x_t−x_{t−1}||^2 + vmaxδ), which enforces ||x_t−x_{t−1}||^2 ≤ vmaxδ. However, the stated speed constraint in the same section is ||x_t−x_{t−1}|| ≤ vmaxδ, with vmaxδ = 4.44 m for the reported settings. As written, the barrier restricts displacements to at most about 2.1 m, contradicting the maximum-distance definition and the graph edge parameter K = ⌈vmaxδ/γ⌉. Since MSR provides the critical initialization (Table VI shows random initialization performs much worse), this inconsistency affects the whole pipeline. Please correct the barrier to use (vmaxδ)^2 or an equivalent norm constraint.
  4. [Section V-A and Eq. (9)] The method requires the positions o_q of all base stations to build the regression matrix C_q in Eq. (9) and to score candidate positions in Eq. (1). The paper never states how the o_q values were obtained for the two live 5G networks (operator data, public databases, survey, or estimation), nor does it quantify sensitivity to BS-position errors. Without this information, the claimed 12–15 m QLE is not interpretable or reproducible, and the comparison against baselines becomes difficult to assess. The paper also does not describe how the ground-truth trajectories used for QLE/TME were recorded. Please add the data-provenance details and a perturbation analysis with respect to o_q errors.
minor comments (6)
  1. [Section II-D, Eq. (3)] In Eq. (3), the first Gaussian factor uses σ_v while the exponent uses σ_{a,v}; the subscript is missing in the normalization term. The accompanying text also calls σ_v a variance, although it appears to denote a standard deviation.
  2. [Section V-C, Table III] The text states that on Dataset II the HRE achieves a QLE of 11.2 m, but Table III reports 14.7 m. Please reconcile the numbers.
  3. [Figure 7 caption] The caption lists the parameters for panels (c) and (d) both as γ(2); one of them should be γ(1), consistent with the text.
  4. [Figures 4 and 8] The legends include a 'DCL' method that is not defined anywhere in the paper and does not appear in Table III; please define it or remove it.
  5. [Section IV-A] The convergence argument for Algorithm 1 says that convergence is ensured because both subproblems are solved optimally. This is incomplete; the monotone increase of the objective and boundedness of the objective should be stated explicitly.
  6. [General] No code or data are provided. Given the implementation ambiguities in Eq. (2), Algorithm 2, and Eq. (16), making code available would substantially improve verifiability.

Circularity Check

0 steps flagged · score 1.0 of 10

No definitional circularity: the recovered trajectory is determined by raw RSS through a fitted multi-BS path-loss likelihood and scored against external GPS ground truth; self-citations are peripheral, and the Eq. (2) Lambert-W domain flaw is a correctness issue outside the circularity rubric.

full rationale

Verdict: no definitional circularity. The target of prediction is the trajectory, evaluated by QLE/TME against externally collected GPS ground truth; neither the fitted propagation parameters Theta nor the mobility parameters Phi are benchmarked as predictions. Algorithm 1 alternates between closed-form least-squares fits of per-BS path-loss parameters (Eq. 10) and HMM decoding of the trajectory (Algorithm 2). This is EM-like mutual dependence, but the recovered path is pinned by the raw multi-BS RSS likelihood (Eq. 1): the per-cell regression can only pass through positions consistent with the observed RSS vector, and the WCL baseline's 85 m QLE shows the RSS data carry genuine position information. Table VI's oracle-parameter ablation (Theta(GT): QLE 8.4 m vs. fitted Theta: 12.6 m) shows the fitted model approaches the ground-truth-parameter bound, so the result is not forced by the fit by construction. Mobility parameters in the fixed model (vavr, eta, vmax) are explicitly declared priors, and the adaptive variant's group speed moments (Eqs. 14-15) are standard EM nuisance parameters. Self-citations [5], [18], [20], [22], [57] appear only in the introduction's survey and are not load-bearing. The review policy requires flagging Eq. (2): sigma-hat^2_v(eta) uses branch W_{-1}, real only for arguments >= -1/e ~ -0.368, yet the paper's own settings (eta=0.05, vmax-vavr=11.7) give -2*pi*(0.05)^2*(11.7)^2 ~ -2.15, so the transition probability is not real-valued for the reported experiments. This is a serious correctness/executability flaw, but it is not circularity, so it does not raise the circularity score.

Assumptions & free parameters 8 free parameters · 7 assumptions · 0 invented entities

The central claim rests on hand-set or fitted parameters, including per-BS path-loss coefficients, mobility priors, graph spacings, and speed-group counts, plus domain assumptions that tower coordinates are known, shadowing is Gaussian and independent, and the road map matches reality. No code or data are released, and a core speed-variance formula is undefined at the stated parameter values.

free parameters (8)
  • Per-BS path-loss parameters alpha_q, beta_q, sigma_q = Not reported; estimated in Eq. (10)
    These are fit to the same RSS and trajectory inside the alternating loop, so the observation model is not independently calibrated.
  • Average speed v_avr = 10.5 m/s (Dataset I), 10.8 m/s (Dataset II)
    Set by hand from road-type priors; the true average speeds are 13.3 and 13.8 m/s, so the prior is deliberately lower.
  • Max speed v_max = 22.2 m/s (80 km/h)
    Chosen from city speed limit; actual measured maxima are 105.7 and 110.5 km/h, exceeding the prior.
  • Maximum-speed probability or density eta = 0.05
    Hand-set to define speed variance via Eq. (2); the formula is undefined at this value, which is a flagged red flag.
  • Number of adaptive speed groups A = 10
    Hand-set; group boundaries come from equal-width bins of the normalized RSS variation.
  • Coarse and fine graph spacings gamma(1), gamma(2) = gamma(1)=300 m, gamma(2)=2 m
    Chosen to balance state count and resolution; sensitivity to gamma(2) is reported.
  • Per-group mobility parameters (v_a,avr, sigma^2_a,v) = Estimated via Eqs. (14)-(15)
    In HREA these are learned from the current trajectory for each of A groups.
  • Interior-point hyperparameters c(0), mu, epsilon1, epsilon2 = Selected per Boyd and Vandenberghe [62]
    Standard settings for the barrier method; not central to the trajectory claim.
assumptions (7)
  • domain assumption RSS follows y = beta_q + alpha_q log10(d) + epsilon with independent Gaussian shadowing per BS (Sec. II-B).
    All observation probabilities derive from this; real shadowing is spatially and temporally correlated, and the reported strongest-beam RSS is not zero-mean.
  • domain assumption Conditional independence of RSS observations across BSs and time slots given position (Eq. (1)).
    Required for the product observation likelihood; ignored correlations could bias path scores.
  • domain assumption BS positions o_q are known exactly (Eq. (9)).
    The distance norms in C_q require tower coordinates; the paper never states how they were obtained or how accurate they are.
  • domain assumption Vehicle speeds follow a Gaussian distribution with given or segment-wise parameters (Sec. II-C, II-D).
    Transition probabilities and the adaptive grouping depend on this model.
  • domain assumption A digital road network graph with nodes every gamma meters and edges up to K*gamma is available (Sec. II-C).
    The map source and coverage are not described; the Viterbi search is only as good as the graph.
  • domain assumption The ground-truth trajectory used for QLE and TME is accurate (Sec. V-A).
    Evaluation assumes the phone-recorded reference path is correct to well below the reported errors.
  • standard math Viterbi decoding on the road graph gives the global optimum of subproblem (7) (Sec. IV-C).
    Holds if graph edges cover all feasible moves and scores are additive log-probabilities; the argmin and argmax inconsistency in Algorithm 2 needs correction.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Trajectory Map-Matching in Urban Road Networks Based on RSS Measurements." pith.science (2026). https://pith.science/paper/R4JVVVGF

@misc{pith2026250201280,
  author       = {Pith},
  title        = {Pith review of: Trajectory Map-Matching in Urban Road Networks Based on RSS Measurements},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/R4JVVVGF}},
  note         = {Machine review of arXiv:2502.01280}
}
read the original abstract

This paper proposes an RSS-based approach to reconstruct vehicle trajectories within a road network, enforcing signal propagation rules and vehicle mobility constraints to mitigate the impact of RSS noise and sparsity. The key challenge lies in leveraging latent spatiotemporal correlations within RSS data while navigating complex road networks. To address this, we develop a Hidden Markov Model (HMM)-based RSS embedding (HRE) technique that employs alternating optimization to infer vehicle trajectories from RSS measurements. This model captures spatiotemporal dependencies while a road graph ensures network compliance. Additionally, we introduce a maximum speed-constrained rough trajectory estimation (MSR) method to guide the optimization process, enabling rapid convergence to a favorable local solution.

Figures

Figures reproduced from arXiv: 2502.01280 by the authors.

Figure 1
Figure 1. Illustration of the relationship between location space and RSS [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. The frequency histogram of real speeds and recovered speeds v.s. the [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. (a) Relationship between trajectory length and the runtime of [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figures from the paper (3 more)
Figure 6
Figure 6. Figure 6: Example procedure of Algorithm 2 with γ (1) = 300, γ(2) = 2. The blue dots represent nodes in the road graph. to all baseline methods and the proposed HRE on Dataset I. Similarly, on Dataset II, it achieves the smallest TME of 0.25% and a QLE of 12.4 meters when compar…
Figure 5
Figure 5. Figure 5: Example trajectory estimation result of our HRE with different missing [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 8
Figure 8. Figure 8: Relationship between noise level and the Performance of the proposed [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

75 extracted references · 55 canonical work pages

  1. [1]

    Cellular localization for autonomous driving: A function pull approach to safety-critical wireless localization,

    R. Whiton, “Cellular localization for autonomous driving: A function pull approach to safety-critical wireless localization,”IEEE Veh. Technol. Mag., vol. 17, no. 4, pp. 28–37, 2022

    work page 2022

  2. [2]

    L3-net: Towards learning based lidar localization for autonomous driving,

    W. Lu, Y . Zhou, G. Wan, S. Hou, and S. Song, “L3-net: Towards learning based lidar localization for autonomous driving,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2019, pp. 6389–6398

    2019

  3. [3]

    Feature detection for vehicle localization in urban environments using a multilayer lidar,

    A. Y . Hata and D. F. Wolf, “Feature detection for vehicle localization in urban environments using a multilayer lidar,” IEEE Trans. Intell. Transp. Syst., vol. 17, no. 2, pp. 420–429, 2015

    work page 2015

  4. [4]

    Vehicle trajectory reconstruction at signalized intersections under connected and automated vehicle environment,

    X. Chen, J. Yin, K. Tang, Y . Tian, and J. Sun, “Vehicle trajectory reconstruction at signalized intersections under connected and automated vehicle environment,” IEEE Trans. Intell. Transp. Syst. , vol. 23, no. 10, pp. 17 986–18 000, 2022

    work page 2022

  5. [5]

    Calibration-free indoor positioning via regional channel tracing,

    Z. Xing and W. Zhao, “Calibration-free indoor positioning via regional channel tracing,” IEEE Internet Things J. , vol. 0, no. 0, pp. 0–0, 2024

    work page 2024

  6. [6]

    Enhancing localization accuracy of mems-ins/gps/in-vehicle sensors integration during gps outages,

    Q. Xu, X. Li, and C.-Y . Chan, “Enhancing localization accuracy of mems-ins/gps/in-vehicle sensors integration during gps outages,” IEEE Trans. Instrum. Meas. , vol. 67, no. 8, pp. 1966–1978, 2018

    work page 1966

  7. [7]

    Hidden markov map matching through noise and sparseness,

    P. Newson and J. Krumm, “Hidden markov map matching through noise and sparseness,” in Proceedings of the ACM SIGSPATIAL International Conference on Advances in Geographic Information Systems , 2009, pp. 336–343

    work page 2009

  8. [8]

    Map- matching for low-sampling-rate gps trajectories,

    Y . Lou, C. Zhang, Y . Zheng, X. Xie, W. Wang, and Y . Huang, “Map- matching for low-sampling-rate gps trajectories,” in Proceedings of the 17th ACM SIGSPATIAL International Conference on Advances in Geographic Information Systems , 2009, pp. 352–361

    work page 2009

Show all 75 references
  1. [9]

    Map matching for low-sampling-rate gps trajectories by exploring real-time moving directions,

    Y .-L. Hsueh and H.-C. Chen, “Map matching for low-sampling-rate gps trajectories by exploring real-time moving directions,” Inf. Sci., vol. 433, pp. 55–69, 2018

    2018

  2. [10]

    Transformer-based map-matching model with limited labeled data using transfer-learning approach,

    Z. Jin, J. Kim, H. Yeo, and S. Choi, “Transformer-based map-matching model with limited labeled data using transfer-learning approach,” Transp. Res. Part C: Emerg. Technol. , vol. 140, p. 103668, 2022

    2022

  3. [11]

    Amm: An adaptive online map matching algorithm,

    H. Hu, S. Qian, J. Ouyang, J. Cao, H. Han, J. Wang, and Y . Chen, “Amm: An adaptive online map matching algorithm,”IEEE Trans. Intell. Transp. Syst., vol. 24, pp. 0–0, 2023

    2023

  4. [12]

    Vehicle trajectory prediction method based on license plate information obtained from video-imaging detectors in urban road environment,

    Z. Zhang, H. Liu, L. Rai, and S. Zhang, “Vehicle trajectory prediction method based on license plate information obtained from video-imaging detectors in urban road environment,” Sensors, vol. 20, no. 5, p. 1258, 2020

    2020

  5. [13]

    Bluetooth data in an urban context: Retrieving vehicle trajectories,

    G. Michau, A. Nantes, A. Bhaskar, E. Chung, P. Abry, and P. Borgnat, “Bluetooth data in an urban context: Retrieving vehicle trajectories,” IEEE Trans. Intell. Transp. Syst. , vol. 18, no. 9, pp. 2377–2386, 2017

    2017

  6. [14]

    A cloud-friendly rfid trajectory clustering algorithm in uncertain environments,

    Y . Wu, H. Shen, and Q. Z. Sheng, “A cloud-friendly rfid trajectory clustering algorithm in uncertain environments,” IEEE Trans. Parallel Distrib. Syst., vol. 26, no. 8, pp. 2075–2088, 2014

    2014

  7. [15]

    Spatio-temporal trajectory estimation based on incomplete wi-fi probe data in urban rail transit network,

    J. Gu, Z. Jiang, Y . Sun, M. Zhou, S. Liao, and J. Chen, “Spatio-temporal trajectory estimation based on incomplete wi-fi probe data in urban rail transit network,” Knowl.-Based Syst., vol. 211, p. 106528, 2021

    2021

  8. [16]

    Vehicle localization utilizing a novel hybrid tdoa-based estimation,

    O. Owen, Z. Pan, and S. Shimamoto, “Vehicle localization utilizing a novel hybrid tdoa-based estimation,” in Proceedings of the IEEE Vehicular Technology Conference, 2022, pp. 1–6

    2022

  9. [17]

    Feasibility study of 5g-based localization for assisted driv- ing,

    J. A. del Peral-Rosado, J. A. L ´opez-Salcedo, S. Kim, and G. Seco- Granados, “Feasibility study of 5g-based localization for assisted driv- ing,” in Proceedings of the International Conference on Localization and GNSS, 2016, pp. 1–6

    2016

  10. [18]

    Block-diagonal guided dbscan clustering,

    Z. Xing and W. Zhao, “Block-diagonal guided dbscan clustering,” IEEE Trans. Knowl. Data Eng. , vol. 0, no. 0, pp. 0–0, 2024

    2024

  11. [19]

    Analysis of knn density estimation,

    P. Zhao and L. Lai, “Analysis of knn density estimation,” IEEE Trans- actions on Information Theory , vol. 68, no. 12, pp. 7971–7995, 2022

    2022

  12. [20]

    Clustering structure identification with ordering graph,

    Z. Xing and W. Zhao, “Clustering structure identification with ordering graph,” 2023. [Online]. Available: https://openreview.net/forum?id= HG0SwOmlaEo

    2023

  13. [21]

    Robust nonparametric regression under poisoning attack,

    P. Zhao and Z. Wan, “Robust nonparametric regression under poisoning attack,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 15, 2024, pp. 17 007–17 015

    2024

  14. [22]

    Block-diagonal structure learning for subspace clustering,

    Z. Xing and W. Zhao, “Block-diagonal structure learning for subspace clustering,” 2023. [Online]. Available: https://openreview.net/forum?id= sxLL8K3E39G

    2023

  15. [23]

    Trip matrix and path flow reconstruction and estimation based on plate scanning and link observations,

    E. Castillo, J. M. Men ´endez, and P. Jim ´enez, “Trip matrix and path flow reconstruction and estimation based on plate scanning and link observations,” Transp. Res. Part B: Methodol. , vol. 42, no. 5, pp. 455– 481, 2008

    2008

  16. [24]

    Pointlie: Locally invertible embedding for point cloud sampling and recovery,

    W. Zhao, X. Yan, J. Gao, R. Zhang, J. Zhang, Z. Li, S. Wu, and S. Cui, “Pointlie: Locally invertible embedding for point cloud sampling and recovery,” in Proceedings of the International Joint Conference on Artificial Intelligence, 2021, p. 0

    2021

  17. [25]

    Wivelo: Fine-grained walking ve- locity estimation for wi-fi passive tracking,

    C. Li, L. Liu, Z. Cao, and M. Zhang, “Wivelo: Fine-grained walking ve- locity estimation for wi-fi passive tracking,” in Proceedings of the IEEE International Conference on Sensing, Communication, and Networking , 2022, pp. 172–180

    2022

  18. [26]

    A huber loss minimization approach to byzantine robust federated learning,

    P. Zhao, F. Yu, and Z. Wan, “A huber loss minimization approach to byzantine robust federated learning,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 38, no. 19, 2024, pp. 21 806– 21 814

    2024

  19. [27]

    Multiple distresses detection for asphalt pavement using improved you only look once algorithm based on convolutional neural network,

    H.-C. Dan, P. Yan, J. Tan, Y . Zhou, and B. Lu, “Multiple distresses detection for asphalt pavement using improved you only look once algorithm based on convolutional neural network,” International Journal of Pavement Engineering, vol. 25, no. 1, p. 2308169, 2024

    2024

  20. [28]

    Research progress of all-fiber optic current transformers in novel power systems: A review,

    Z. Li, J. Cui, H. Chen, H. Lu, F. Zhou, P. R. F. Rocha, and C. Yang, “Research progress of all-fiber optic current transformers in novel power systems: A review,” Microwave and Optical Technology Letters, vol. 67, no. 1, p. e70061, 2025

    2025

  21. [29]

    Using decision tree classification and principal component analysis to predict ethnicity based on individual characteristics: A case study of assam and bhutan ethnicities,

    T. Zhang, X. Zhang, X. Liu, Z. Guo, and Y . Tian, “Using decision tree classification and principal component analysis to predict ethnicity based on individual characteristics: A case study of assam and bhutan ethnicities,” Journal of Software Engineering and Applications , vo...

    2024

  22. [30]

    You can even annotate text with voice: Transcription-only-supervised text spotting,

    J. Tang, S. Qiao, B. Cui, Y . Ma, S. Zhang, and D. Kanoulas, “You can even annotate text with voice: Transcription-only-supervised text spotting,” in Proceedings of the 30th ACM International Conference on Multimedia , ser. MM ’22. New York, NY , USA: Association for Computing...

    2022

  23. [31]

    Deep reinforcement learning-based obstacle avoidance for robot movement in warehouse environments,

    K. Li, J. Chen, D. Yu, T. Dajun, X. Qiu, J. Lian, R. Ji, S. Zhang, Z. Wan, B. Sun et al. , “Deep reinforcement learning-based obstacle avoidance for robot movement in warehouse environments,” in 2024 IEEE 6th International Conference on Civil Aviation Safety and Information Te...

    2024

  24. [32]

    Teg-di: Dynamic incentive model for federated learning based on tripartite evolutionary game,

    J. Xu, Y . Zhao, X. Li, L. Zhou, K. Zhu, X. Xu, Q. Duan, and R. Zhang, “Teg-di: Dynamic incentive model for federated learning based on tripartite evolutionary game,” Neurocomputing, p. 129259, 2024. 14

    2024

  25. [33]

    Minimax optimal q learning with nearest neigh- bors,

    P. Zhao and L. Lai, “Minimax optimal q learning with nearest neigh- bors,” IEEE Transactions on Information Theory , 2024

    2024

  26. [34]

    Bearing fault diagnostic framework under unknown working conditions based on condition- guided diffusion model,

    Z. Guo, L. Xu, Y . Zheng, J. Xie, and T. Wang, “Bearing fault diagnostic framework under unknown working conditions based on condition- guided diffusion model,” Measurement, vol. 242, p. 115951, 2025

    2025

  27. [35]

    Prediction model of measurement errors in current transformers based on deep learning,

    Z.-h. Li, J.-x. Cui, H.-p. Lu, F. Zhou, Y .-l. Diao, and Z.-x. Li, “Prediction model of measurement errors in current transformers based on deep learning,” Review of Scientific Instruments, vol. 95, no. 4, p. 044704, 04 2024

    2024

  28. [36]

    A two-stage federated learning method for personalization via selective collaboration,

    J. Xu, L. Zhou, Y . Zhao, X. Li, K. Zhu, X. Xu, Q. Duan, and R. Zhang, “A two-stage federated learning method for personalization via selective collaboration,” Computer Communications, p. 108053, 2025

    2025

  29. [37]

    Optimizing automated picking systems in ware- house robots using machine learning,

    K. Li, J. Wang, X. Wu, X. Peng, R. Chang, X. Deng, Y . Kang, Y . Yang, F. Ni, and B. Hong, “Optimizing automated picking systems in ware- house robots using machine learning,” arXiv preprint arXiv:2408.16633, 2024

    2024

  30. [38]

    Machine learning optimizes the efficiency of picking and packing in automated warehouse robot systems,

    D. Yu, L. Liu, S. Wu, K. Li, C. Wang, J. Xie, R. Chang, Y . Wang, Z. Wang, and R. Ji, “Machine learning optimizes the efficiency of picking and packing in automated warehouse robot systems,” in 2024 In- ternational Conference on Computer Engineering, Network and Digital Commun...

    2024

  31. [39]

    Detection of ai deepfake and fraud in online payments using gan-based models,

    Z. Ke, S. Zhou, Y . Zhou, C. H. Chang, and R. Zhang, “Detection of ai deepfake and fraud in online payments using gan-based models,” arXiv preprint arXiv:2501.07033, 2025

    2025

  32. [40]

    Identifying money laundering risks in digital asset transactions based on ai algorithms,

    Q. Yu, Z. Ke, G. Xiong, Y . Cheng, and X. Guo, “Identifying money laundering risks in digital asset transactions based on ai algorithms,” 2025

    2025

  33. [41]

    Imagpose: A unified conditional framework for pose-guided person generation,

    F. Shen and J. Tang, “Imagpose: A unified conditional framework for pose-guided person generation,” in The Thirty-eighth Annual Conference on Neural Information Processing Systems , 2024

    2024

  34. [42]

    Imagdressing-v1: Customizable virtual dressing,

    F. Shen, X. Jiang, X. He, H. Ye, C. Wang, X. Du, Z. Li, and J. Tang, “Imagdressing-v1: Customizable virtual dressing,” arXiv preprint arXiv:2407.12705, 2024

    2024 arXiv

  35. [43]

    Advancing pose-guided image synthesis with progressive conditional diffusion models,

    F. Shen, H. Ye, J. Zhang, C. Wang, X. Han, and Y . Wei, “Advancing pose-guided image synthesis with progressive conditional diffusion models,” in The Twelfth International Conference on Learning Repre- sentations

  36. [44]

    Boosting consistency in story visualization with rich-contextual condi- tional diffusion models,

    F. Shen, H. Ye, S. Liu, J. Zhang, C. Wang, X. Han, and W. Yang, “Boosting consistency in story visualization with rich-contextual condi- tional diffusion models,” arXiv preprint arXiv:2407.02482 , 2024

    2024 arXiv

  37. [45]

    Pedestrian-specific bipartite-aware similarity learning for text-based person retrieval,

    F. Shen, X. Shu, X. Du, and J. Tang, “Pedestrian-specific bipartite-aware similarity learning for text-based person retrieval,” in Proceedings of the 31th ACM International Conference on Multimedia , 2023

    2023

  38. [46]

    Covid-19 localization and recognition on chest radiographs based on yolov5 and efficientnet,

    T. Zhang, B. Zhang, F. Zhao, and S. Zhang, “Covid-19 localization and recognition on chest radiographs based on yolov5 and efficientnet,” in 2022 7th International Conference on Intelligent Computing and Signal Processing (ICSP). IEEE, 2022, pp. 1827–1830

    2022

  39. [47]

    A mixed-heuristic quantum- inspired simplified swarm optimization algorithm for scheduling of real- time tasks in the multiprocessor system,

    P.-C. Su, S.-Y . Tan, Z. Liu, and W.-C. Yeh, “A mixed-heuristic quantum- inspired simplified swarm optimization algorithm for scheduling of real- time tasks in the multiprocessor system,” Applied Soft Computing , vol. 131, p. 109807, 2022

    2022

  40. [48]

    Accurate prediction of temperature indicators in eastern china using a multi-scale cnn-lstm-attention model,

    J. Shen, W. Wu, and Q. Xu, “Accurate prediction of temperature indicators in eastern china using a multi-scale cnn-lstm-attention model,” arXiv preprint arXiv:2412.07997 , 2024

    2024 arXiv

  41. [49]

    Deep learning-based anomaly detection and log analysis for computer networks,

    S. Wang, R. Jiang, Z. Wang, and Y . Zhou, “Deep learning-based anomaly detection and log analysis for computer networks,” arXiv preprint arXiv:2407.05639, 2024

    2024 arXiv

  42. [50]

    Evaluation of asphalt pavement texture using multiview stereo reconstruction based on deep learning,

    H.-C. Dan, B. Lu, and M. Li, “Evaluation of asphalt pavement texture using multiview stereo reconstruction based on deep learning,” Con- struction and Building Materials , vol. 412, p. 134837, 2024

    2024

  43. [51]

    Energy-efficiency for smartphones using interaction link prediction in mobile cloud computing,

    J. Xu, C. Guan, and X. Xu, “Energy-efficiency for smartphones using interaction link prediction in mobile cloud computing,” in Computer Supported Cooperative Work and Social Computing: 13th CCF Confer- ence, ChineseCSCW 2018, Guilin, China, August 18–19, 2018, Revised Selected...

    2018

  44. [52]

    Dmaf: data-model anti-forgetting for federated incremental learning,

    K. Zhu, J. Xu, L. Zhou, X. Li, Y . Zhao, X. Xu, and S. Li, “Dmaf: data-model anti-forgetting for federated incremental learning,” Cluster Computing, vol. 28, no. 1, p. 30, 2025

    2025

  45. [53]

    Few could be better than all: Feature sampling and grouping for scene text detection,

    J. Tang, W. Zhang, H. Liu, M. Yang, B. Jiang, G. Hu, and X. Bai, “Few could be better than all: Feature sampling and grouping for scene text detection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2022, pp. 4563–4572

    2022

  46. [54]

    Spts v2: single-point scene text spotting,

    Y . Liu, J. Zhang, D. Peng, M. Huang, X. Wang, J. Tang, C. Huang, D. Lin, C. Shen, X. Bai et al., “Spts v2: single-point scene text spotting,” IEEE Transactions on Pattern Analysis and Machine Intelligence , 2023

    2023

  47. [55]

    Research on reinforcement learning based warehouse robot navigation algorithm in complex warehouse layout,

    K. Li, L. Liu, J. Chen, D. Yu, X. Zhou, M. Li, C. Wang, and Z. Li, “Research on reinforcement learning based warehouse robot navigation algorithm in complex warehouse layout,” in 2024 6th International Con- ference on Artificial Intelligence and Computer Applications (ICAICA) ...

    2024

  48. [56]

    Image-driven prediction system: Automatic extraction of aggregate gradation of pavement core samples integrating deep learning and interactive image processing framework,

    H.-C. Dan, Z. Huang, B. Lu, and M. Li, “Image-driven prediction system: Automatic extraction of aggregate gradation of pavement core samples integrating deep learning and interactive image processing framework,” Construction and Building Materials , vol. 453, p. 139056, 2024

    2024

  49. [57]

    Unsupervised action segmentation via fast learning of semantically consistent actoms,

    Z. Xing and W. Zhao, “Unsupervised action segmentation via fast learning of semantically consistent actoms,” in Proceedings of the AAAI Conference on Artificial Intelligence , 2024, pp. 6270–6278

    2024

  50. [58]

    Optimal boxes: boosting end-to-end scene text recognition by adjusting annotated bounding boxes via reinforcement learning,

    J. Tang, W. Qian, L. Song, X. Dong, L. Li, and X. Bai, “Optimal boxes: boosting end-to-end scene text recognition by adjusting annotated bounding boxes via reinforcement learning,” in European Conference on Computer Vision. Springer, 2022, pp. 233–248

    2022

  51. [59]

    Multi-modal in-context learning makes an ego-evolving scene text recognizer,

    Z. Zhao, J. Tang, C. Lin, B. Wu, C. Huang, H. Liu, X. Tan, Z. Zhang, and Y . Xie, “Multi-modal in-context learning makes an ego-evolving scene text recognizer,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, pp. 15 567–15 576

    2024

  52. [60]

    Character recognition competition for street view shop signs,

    J. Tang, W. Du, B. Wang, W. Zhou, S. Mei, T. Xue, X. Xu, and H. Zhang, “Character recognition competition for street view shop signs,” National Science Review, vol. 10, no. 6, p. nwad141, 2023

    2023

  53. [61]

    Spot: Structure patching and overlap tweaking for effective pipelining in privacy-preserving mlaas with tiny clients,

    X. Xu, Q. Zhang, R. Ning, C. Xin, and H. Wu, “Spot: Structure patching and overlap tweaking for effective pipelining in privacy-preserving mlaas with tiny clients,” in 2024 IEEE 44th International Conference on Distributed Computing Systems (ICDCS) . IEEE Computer Society, 202...

    2024

  54. [62]

    S. P. Boyd and L. Vandenberghe, Convex Optimization . Cambridge University Press, 2004

    2004

  55. [63]

    Devel- opment of origin–destination matrices using mobile phone call data,

    M. S. Iqbal, C. F. Choudhury, P. Wang, and M. C. Gonz ´alez, “Devel- opment of origin–destination matrices using mobile phone call data,” Transp. Res. Part C: Emerg. Technol. , vol. 40, pp. 63–74, 2014

    2014

  56. [64]

    Accurate analysis of weighted centroid localization,

    K. Magowe, A. Giorgetti, S. Kandeepan, and X. Yu, “Accurate analysis of weighted centroid localization,” IEEE Trans. Cogn. Commun. Netw. , vol. 5, no. 1, pp. 153–164, 2018

    2018

  57. [65]

    Docpedia: Unleashing the power of large multimodal model in the frequency domain for versatile document understanding,

    H. Feng, Q. Liu, H. Liu, J. Tang, W. Zhou, H. Li, and C. Huang, “Docpedia: Unleashing the power of large multimodal model in the frequency domain for versatile document understanding,” Science China Information Sciences, vol. 67, no. 12, pp. 1–14, 2024

    2024

  58. [66]

    Harmonizing visual text comprehension and generation,

    Z. Zhao, J. Tang, B. Wu, C. Lin, S. Wei, H. Liu, X. Tan, Z. Zhang, C. Huang, and Y . Xie, “Harmonizing visual text comprehension and generation,” arXiv preprint arXiv:2407.16364 , 2024

    2024 arXiv

  59. [67]

    Pargo: Bridging vision-language with partial and global views,

    A.-L. Wang, B. Shan, W. Shi, K.-Y . Lin, X. Fei, G. Tang, L. Liao, J. Tang, C. Huang, and W.-S. Zheng, “Pargo: Bridging vision-language with partial and global views,” arXiv preprint arXiv:2408.12928 , 2024

    2024 arXiv

  60. [68]

    Attentive eraser: Unleashing diffusion model’s object removal potential via self-attention redirection guidance,

    W. Sun, B. Cui, J. Tang, and X.-M. Dong, “Attentive eraser: Unleashing diffusion model’s object removal potential via self-attention redirection guidance,” arXiv preprint arXiv:2412.12974 , 2024

    2024

  61. [69]

    A bounding box is worth one token: Interleaving layout and text in a large language model for document understanding,

    J. Lu, H. Yu, Y . Wang, Y . Ye, J. Tang, Z. Yang, B. Wu, Q. Liu, H. Feng, H. Wang et al., “A bounding box is worth one token: Interleaving layout and text in a large language model for document understanding,” arXiv preprint arXiv:2407.01976, 2024

    2024 arXiv

  62. [70]

    Tabpedia: Towards comprehensive visual table understanding with concept synergy,

    W. Zhao, H. Feng, Q. Liu, J. Tang, S. Wei, B. Wu, L. Liao, Y . Ye, H. Liu, W. Zhou et al., “Tabpedia: Towards comprehensive visual table understanding with concept synergy,” arXiv preprint arXiv:2406.01326, 2024

    2024 arXiv

  63. [71]

    Mtvqa: Benchmarking multilingual text-centric visual question answering,

    J. Tang, Q. Liu, Y . Ye, J. Lu, S. Wei, C. Lin, W. Li, M. F. F. B. Mahmood, H. Feng, Z. Zhao et al., “Mtvqa: Benchmarking multilingual text-centric visual question answering,” arXiv preprint arXiv:2405.11985 , 2024

    2024 arXiv

  64. [72]

    Textsquare: Scaling up text-centric visual instruction tuning,

    J. Tang, C. Lin, Z. Zhao, S. Wei, B. Wu, Q. Liu, H. Feng, Y . Li, S. Wang, L. Liao et al. , “Textsquare: Scaling up text-centric visual instruction tuning,” arXiv preprint arXiv:2404.12803 , 2024

    2024 arXiv

  65. [73]

    Mctbench: Multimodal cognition towards text- rich visual scenes benchmark,

    B. Shan, X. Fei, W. Shi, A.-L. Wang, G. Tang, L. Liao, J. Tang, X. Bai, and C. Huang, “Mctbench: Multimodal cognition towards text- rich visual scenes benchmark,” arXiv preprint arXiv:2410.11538 , 2024

    2024 arXiv

  66. [74]

    Unidoc: A universal large multimodal model for simultaneous text detection, recognition, spotting and understanding,

    H. Feng, Z. Wang, J. Tang, J. Lu, W. Zhou, H. Li, and C. Huang, “Unidoc: A universal large multimodal model for simultaneous text detection, recognition, spotting and understanding,” arXiv preprint arXiv:2308.11592, 2023

    2023 arXiv

  67. [75]

    Comet: A communication-efficient and performant approximaion for private trans- former inference,

    X. Xu, Q. Zhang, R. Ning, C. Xin, and H. Wu, “Comet: A communication-efficient and performant approximaion for private trans- former inference,” arXiv preprint arXiv:2405.17485 , 2024

    2024 arXiv

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.