Pith. sign in

REVIEW 5 major objections 4 minor 34 references

CiUAV: A Multi-Task 3D Indoor Localization System for UAVs based on Channel State Information

T0 review · 5 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a passive, low-cost ESP32 channel-state-information system can localize an indoor UAV in 3D with an LMSE error of 0.2629 m, using dynamic AGC compensation and a multi-task Sensor-in-Sample model.

desk verdict The hardware and AGC compensation idea are plausible, but the headline accuracy is uninterpretable because the paper swaps RMSE and MSE and never specifies the SiS model. read the letter →

arxiv 2505.21216 v1 pith:UORM6KSQ submitted 2025-05-27 eess.SP

Cunyi Yin , Chenwei Wang , Jing Chen , Hao Jiang , Xiren Miao , Shaocong Zheng Zhenghua Chen Senior , Hong Yan This is my paper
classification eess.SP
keywords indoorUAVlocalizationchannelstateinformationESP32passivesniffingautomaticgaincontrolmulti-tasklearningSensor-in-Sample3Dpositioning
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 tries to establish that accurate 3D indoor UAV localization is possible without expensive tracking hardware and without adding any equipment to the drone. It builds a system around cheap ESP32 WiFi sensors that passively sniff the UAV's normal WiFi traffic, extracts channel state information (CSI), and feeds that signal data into a neural network that predicts the drone's 3D coordinates. The central result is a localization mean squared error of 0.2629 m in a 5 m by 5 m test space, with more than 80% of errors below 0.25 m when using the best multi-task model. This matters because indoor UAVs operate in GPS-denied settings where cost, payload limits, and robustness to missing sensors or limited training data are practical blockers for deployment.

What carries the argument

The central mechanism is the combination of two components. First, the dynamic AGC compensation algorithm: the ESP32 exposes real-time AGC gain values in decibels, which the system converts to a linear scaling factor $\rho = 1/10^{\text{AGC}_{\text{gain}}/20}$ and applies to each subcarrier's real and imaginary parts, undoing the amplitude distortion that otherwise masks position-dependent CSI features. Second, the Sensor-in-Sample (SiS) multi-task network: a feature extractor $F_e$ shared across tasks maps CSI inputs $X \in \mathbb{R}^{N\times S\times f}$ to features $H$, and a regression module $F_r$ maps $H$ to 3D coordinates per sensor, with total loss $L_{\text{total}} = L_{\text{pred}} + \lambda_s \|w_s\|_1 + \lambda_v L_{\text{sam}}$ where the sparsity term prunes uninformative sensors and the sample-weight term emphasizes high-value training samples. This jointly trained structure is what allows the model to keep working when some sensors are missing or only a fraction of the training data is available.

What would settle it

Record the drone's trajectory with an independent high-precision reference such as a motion-capture system while running CiUAV; if the Aruco/ToF reference itself has errors of tens of centimeters, or if label timestamps jitter relative to CSI frames, the reported 0.26 m LMSE would not reflect the localization system's true accuracy. A second check is to repeat the ablation with the DAC module disabled and with all CSI amplitudes multiplied by a constant, to confirm the reported 5.7 cm degradation is caused by AGC compensation rather than by label noise.

Watch

Extended reading notes

Core claim

The authors propose and demonstrate CiUAV, a complete CSI-based passive 3D localization pipeline for indoor UAVs. The system uses ESP32-S3 sensors that send ICMP echo requests to the UAV and parse CSI from the returned packets, so no onboard localization hardware or active communication link is needed. A dynamic automatic gain control (AGC) compensation algorithm reads real-time hardware gain values, converts them to linear scaling factors, and rescales each CSI subcarrier to undo amplitude distortion introduced by the AGC circuit. The compensated CSI is then processed by Sensor-in-Sample (SiS), a multi-task network with a shared feature extractor that jointly trains across different sensor configurations and training sample sizes. In experiments with three ceiling-mounted sensors over a 5 m by 5 m flight area, the MobileNet-based SiS model achieves an LMSE of 0.2629 m and an R2 of 0.8559, outperforming a single-task baseline (LMSE 1.6947 m) and the compared CSI localization baselines. Ablation results show that removing the AGC compensation raises LMSE to 0.3451 m, while adding the outlier-handling filter further improves it to the reported 0.2629 m; reducing training data to 25% or dropping to a two-sensor configuration degrades performance gracefully rather than breaking it.

Load-bearing premise

The measured 0.26 m error assumes the Aruco-marker and ToF positions used as ground truth are accurate to well below that error and are time-synchronized with each CSI frame; the paper does not report the accuracy or synchronization of this reference.

Editorial extensions

If this is right

  • If the reported accuracy holds, a multi-UAV indoor deployment could use commodity ESP32 hardware as passive beacons and achieve submeter 3D localization without burdening the drone's payload.
  • The AGC compensation step is load-bearing: removing it degrades the localization error by roughly 5.7 cm, implying that any CSI-based fingerprinting system using gain-controlled WiFi hardware should account for automatic gain in the preprocessing stage.
  • The multi-task design directly addresses sensor failure and limited datasets, meaning the system could remain usable in real deployments where one sensor goes offline or data collection time is short.
  • Passive sniffing via ICMP replies implies localization can run during normal WiFi communication of the UAV, without dedicated ranging frames or cooperation from the flight controller.

Reading between the lines

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

  • Editorial extension: if this result transfers across room sizes and layouts, the same passive-ESP32-plus-AGC pipeline could be applied to other WiFi-emitting mobile robots, not just UAVs, because the method does not depend on drone-specific control signals.
  • Editorial extension: the SiS multi-task training strategy is a generic recipe for CSI fingerprinting with heterogeneous sensor arrays, and could be tested on human activity recognition or device-free gesture localization where sensors occasionally drop out.
  • Editorial extension: a natural stress test would be moving the sensors or changing the flight area after training; the paper does not report cross-environment generalization, so whether the 0.26 m figure survives environmental change remains open.
  • Editorial extension: the AGC compensation depends on the ESP32's ability to expose real-time gain values; applying the same idea to older CSI hardware such as Intel 5300 would require estimating AGC gain indirectly, and the accuracy gap that introduces is untested.
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

5 major / 4 minor

Summary. The manuscript proposes CiUAV, a 3D indoor localization system for UAVs based on CSI collected by low-cost ESP32-S3 sensors. The system passively monitors WiFi packets from a small UAV and uses three contributions: a dynamic AGC compensation algorithm that reads hardware-level AGC gains and rescales CSI amplitudes, a Hampel/outlier processing module, and a multi-task network called Sensor-in-Sample (SiS) that is claimed to maintain accuracy when some sensors are missing or training data are limited. Experiments report a localization error of 0.2629 m (called 'LMSE'), with ablation and robustness studies over data-processing configurations, training-sample proportions, and sensor configurations.

Significance. If the reported accuracy and robustness were validated, the work would be a valuable low-cost, device-free 3D localization approach for small UAVs: it uses commodity ESP32 hardware, exploits directly readable AGC gain information, and addresses realistic sensor/data scarcity. The paper also reports a substantial collected dataset (77,000 training and 33,000 test samples) and compares against two recent CSI localization baselines. As presented, however, the central numerical claim is obscured by a systematic metric confusion, the SiS regularization terms are not connected to the forward model and can be trivially zeroed, the ground-truth labeling is not validated, and at least one key formula is dimensionally inconsistent. The contribution therefore cannot be assessed without substantial correction and additional reporting.

major comments (5)
  1. [§IV.B–IV.E, Tables I–II, Eq. (7)] In every row of Tables I and II, the first column equals the square root of the second column (e.g., 0.5127^2 = 0.2629), so the first column is an RMSE, not an MAE, and the second column is an MSE, not a length in meters. Section IV.D then swaps the labels, stating that the MAE increases from 0.2629 m (100%) to 0.3247 m and the LMSE increases from 0.5127 m to 0.5698 m. The abstract's 'LMSE localization error of 0.2629 m' is therefore dimensionally an area; if the MSE is 0.2629 m^2, the corresponding RMS error is 0.5127 m, a factor-of-two difference from the headline. Section IV.B's claim that 'more than 80% of errors are within 0.25 m' also cannot be reconciled with the reported RMSE/MSE without the full empirical error distribution. Because no code, data, or raw prediction files are provided, this is not merely a typographical issue: the paper never fixes which error statistic is the headline result.
  2. [§III.D, Eqs. (8)–(10)] The regularization terms L_sor = λ_s ||w_s||_1 and L_sam = (1/S) Σ_s ||v ⊙ (Ŷ−Y)||_2^2 are added to L_pred, but neither w_s nor v appears in the forward computation H = F_e(X), Ŷ = F_r(H), or in L_pred. As written, the loss can be minimized by setting w_s = 0 and v = 0, which removes both regularization terms without changing the predictions or the prediction loss. No non-negativity constraint, norm constraint, or explicit gating connection is stated. Thus the claimed sensor-sparsity and sample-selection mechanisms are vacuous, and the robustness results in Sections IV.D and IV.E cannot be attributed to them.
  3. [§III.B and §IV.A] The ground-truth labels are produced by Aruco-marker detection for the (x,y) coordinates and by a ToF sensor for the z coordinate, but the paper reports no calibration or accuracy for this reference and no synchronization procedure between the 50 Hz CSI frames and the labels. If the Aruco/ToF reference has errors on the order of tens of centimeters, or if the label–CSI alignment jitters, the claimed errors of 0.26–0.51 m would be dominated by label noise. The paper must report the reference accuracy, the timestamp alignment method, and ideally an independent validation of the ground-truth pipeline.
  4. [§III.C, Eq. (4)] Equation (4) as typeset reads |H̃(k)| = sqrt(H̃_real(k)^2 + H̃_imag(k)^2) H(k), with an extra factor H(k) after the square root. If H(k) denotes the original complex CSI, this equality implies |H̃(k)| = |H̃(k)| · H(k), which is only true if H(k) = 1. The compensated amplitude should be the square-root expression alone, or the root of the compensated components. This formula needs correction, and the authors should state whether the extra factor was actually used in the experiments.
  5. [§III.D and §IV.E] The ground-truth position of the UAV is a single 3D point, but the model is defined with Y ∈ R^{N×S×D} and Ŷ ∈ R^{N×S×D}, i.e., per-sensor predicted positions, and the loss in Eq. (7) is averaged over the sensor dimension. The paper never states how the S per-sensor predictions are combined to produce the final position used in the reported errors in Tables I and II. This is especially important in the sensor-reduction experiments, where S varies across configurations; without a stated aggregation rule, the evaluation procedure is undefined.
minor comments (4)
  1. [§IV.C] The quantitative ablation statements do not match Table I: the text says Hampel+outlier reduces MAE and LMSE by 2.7 cm and 3.2 cm compared with DAC alone, but the table differences are 0.0239 and 0.0250, respectively; the statement that R^2 decreases by 3 cm also uses a length unit for a dimensionless quantity. Please reconcile the text with the table.
  2. [§III.B] The system is described as 'passive sniffing,' but the CSI sensors actively send ICMP Echo Requests and the UAV must respond with WiFi packets; the terms 'passive' and 'device-free' should be qualified to avoid overclaiming.
  3. [§IV.B] The comparison with LocLite and SSLUL is presented only through CDF curves; please provide the numerical metrics and describe how these baselines were adapted to 3D localization and to ESP32 CSI data.
  4. [§IV.A] The metric 'LMSE' is never defined in Section IV.A, and the model configuration says the input contains N = 10,000 samples while the training set is reported as 77,000 samples; please clarify the relationship between these numbers.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: test-set evaluation and hardware-derived AGC compensation are self-contained; metric inconsistencies are reporting defects, not circular reasoning.

full rationale

The claimed derivation chain is self-contained. The headline localization result is obtained by training the SiS model on a training set ('The training set comprised 77,000 samples, while the test set contained 33,000') and evaluating on a separate dataset collected 'exclusively for testing purposes,' so the 0.2629 m / 0.5127 m figures are test-set measurements, not fitted parameters renamed as predictions. The AGC compensation step is not fitted to localization labels: Eq. (2) defines the scaling factor from the hardware-reported AGC gain in dB (rho = 1/10^{AGCgain/20}), and the description states the ESP32-S3 offers real-time access to AGC gain values, so the correction is derived from hardware state rather than from the target positions. The SiS multi-task loss (Eqs. 7-10) trains on x,y,z labels, but this is standard supervised training; the claimed robustness to sensor/sample reduction is evaluated under different sensor configurations and training proportions, and it does not reduce to the training objective itself. The self-citations ([3], [5], [6], [9], [26]) are background references on indoor localization, remote sensing, and CSI-based estimation; none is invoked as the justification for the central accuracy claim or as a uniqueness theorem, so they are not load-bearing. The paper does have an internal metric inconsistency: Tables I and II list 'MAE' values that are the square roots of the 'LMSE' values (e.g., 0.5127^2 = 0.2629), and Section IV.D swaps the two labels. That is a serious reporting/reproducibility defect and makes the headline 'LMSE localization error of 0.2629 m' dimensionally ambiguous, but it does not make the result identical to an input by construction. The absence of code, data, and ground-truth synchronization details further limits external verification, but these are correctness/reproducibility concerns, not circularity. Therefore the circularity score is 0.

Assumptions & free parameters 4 free parameters · 4 assumptions · 3 invented entities

The ledger shows that the central accuracy claim rests on unverified ground-truth accuracy, an amplitude-only CSI assumption, and two under-specified regularization constructs (w_s, v). The contribution is mostly the dataset and system integration, not a new derivation. No new physical entities are introduced; the invented items are model mechanisms that lack specification.

free parameters (4)
  • Regularization coefficient lambda_s (sensor sparsity) = 0.01
    Chosen by hand in Section IV.A; no sensitivity analysis is provided, and it controls the sensor sparsity penalty.
  • Regularization coefficient lambda_v (sample selection) = 0.1
    Chosen by hand in Section IV.A; it controls the sample-weighting penalty.
  • Feature dimension f_h = 128
    Set in Section IV.A; no ablation or justification is given.
  • Learning rate, batch size, epochs = 1e-4, 32, 400
    Training hyperparameters in Section IV.A; standard choices but no tuning procedure is reported.
assumptions (4)
  • domain assumption CSI amplitude after AGC compensation contains sufficient information to regress 3D drone position
    The entire localization approach relies on amplitude-only CSI features (Section III.D); no phase features or propagation model is used.
  • domain assumption Aruco marker detection and ToF height provide ground-truth positions accurate to well below the claimed error
    Labels are derived from these sensors (Section III.B), but their accuracy and synchronization are not reported.
  • ad hoc to paper Joint training on varying sensor configurations and sample sizes is implemented as claimed
    Section III.D describes multi-task joint training, but no task construction, data sampling schedule, or multi-task loss weighting is specified; the equations shown are a single regression loss with two regularizers.
  • domain assumption The dB-to-linear gain conversion in Eq. (2) fully compensates AGC amplitude distortion
    Eqs. (1)-(3) assume the only AGC effect is a per-symbol linear amplitude scaling; nonlinearities, time-varying gain during a packet, and phase effects are not considered.
invented entities (3)
  • Sensor weight vector w_s
    purpose: Intended to enforce sensor sparsity in the SiS loss (Eq. 8)
    The paper never defines how w_s is computed, what it multiplies, or how it is constrained; it is an ungrounded modeling construct.
  • Sample weight vector v
    purpose: Intended to prioritize high-value training samples in Eq. (9)
    No update rule, constraint, or use in inference is given; without a defined mechanism the sample-selection result cannot be reproduced.
  • Sensor-in-Sample (SiS) multi-task framework
    purpose: Claimed to enable joint training across sensor configurations and sample sizes
    The framework is named but its task structure, task balancing, and shared or private layer configuration are not specified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CiUAV: A Multi-Task 3D Indoor Localization System for UAVs based on Channel State Information." pith.science (2026). https://pith.science/paper/UORM6KSQ

@misc{pith2026250521216,
  author       = {Pith},
  title        = {Pith review of: CiUAV: A Multi-Task 3D Indoor Localization System for UAVs based on Channel State Information},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UORM6KSQ}},
  note         = {Machine review of arXiv:2505.21216}
}
read the original abstract

Accurate indoor positioning for unmanned aerial vehicles (UAVs) is critical for logistics, surveillance, and emergency response applications, particularly in GPS-denied environments. Existing indoor localization methods, including optical tracking, ultra-wideband, and Bluetooth-based systems, face cost, accuracy, and robustness trade-offs, limiting their practicality for UAV navigation. This paper proposes CiUAV, a novel 3D indoor localization system designed for UAVs, leveraging channel state information (CSI) obtained from low-cost ESP32 IoT-based sensors. The system incorporates a dynamic automatic gain control (AGC) compensation algorithm to mitigate noise and stabilize CSI signals, significantly enhancing the robustness of the measurement. Additionally, a multi-task 3D localization model, Sensor-in-Sample (SiS), is introduced to enhance system robustness by addressing challenges related to incomplete sensor data and limited training samples. SiS achieves this by joint training with varying sensor configurations and sample sizes, ensuring reliable performance even in resource-constrained scenarios. Experiment results demonstrate that CiUAV achieves a LMSE localization error of 0.2629 m in a 3D space, achieving good accuracy and robustness. The proposed system provides a cost-effective and scalable solution, demonstrating its usefulness for UAV applications in resource-constrained indoor environments.

Figures

Figures reproduced from arXiv: 2505.21216 by the authors.

Figure 1
Figure 1. Overview of the CiUAV framework, comprising data acquisition, dynamic AGC compensation, model training, and 3D localization. CSI sensors [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The layout of the experiment environment and Planar graph. Passive [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. CDF curves of localization errors for the multi-task 3D localization [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: Localization error distributions for different sensor configurations. The [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

34 extracted references · 33 canonical work pages

  1. [1]

    A comprehensive review of scab disease detection on rosaceae family fruits via uav imagery,

    Z. A. Ali, C. Yang, A. Israr, and Q. Zhu, “A comprehensive review of scab disease detection on rosaceae family fruits via uav imagery,” Drones, vol. 7, no. 2, p. 97, 2023

    work page 2023

  2. [2]

    I am not afraid of the gps jammer: resilient navigation via signals of opportunity in gps- denied environments,

    Z. M. Kassas, J. Khalife, A. A. Abdallah, and C. Lee, “I am not afraid of the gps jammer: resilient navigation via signals of opportunity in gps- denied environments,” IEEE Aerospace and Electronic Systems Magazine, vol. 37, no. 7, pp. 4–19, 2022

    work page 2022

  3. [3]

    Overview of approaches for device het- erogeneity management during indoor localization,

    C. Yin, H. Jiang, and J. Chen, “Overview of approaches for device het- erogeneity management during indoor localization,” in Machine Learning for Indoor Localization and Navigation . Springer, 2023, pp. 259–282

    work page 2023

  4. [4]

    Current status and future trends of meter-level indoor positioning technology: A review,

    L. Qi, Y . Liu, Y . Yu, L. Chen, and R. Chen, “Current status and future trends of meter-level indoor positioning technology: A review,” Remote Sensing, vol. 16, no. 2, p. 398, 2024

    work page 2024

  5. [5]

    Device-free human activity recognition with low-resolution infrared array sensor using long short-term memory neural network,

    C. Yin, J. Chen, X. Miao, H. Jiang, and D. Chen, “Device-free human activity recognition with low-resolution infrared array sensor using long short-term memory neural network,” Sensors, vol. 21, no. 10, p. 3551, 2021

    work page 2021

  6. [6]

    Hu- man activity recognition with low-resolution infrared array sensor using semi-supervised cross-domain neural networks for indoor environment,

    C. Yin, X. Miao, J. Chen, H. Jiang, D. Chen, Y . Tong, and S. Zheng, “Hu- man activity recognition with low-resolution infrared array sensor using semi-supervised cross-domain neural networks for indoor environment,” IEEE Internet of Things Journal, vol. 10, no. 13, pp. 11 761–11 772, 2023

    work page 2023

  7. [7]

    A survey of indoor location technologies, techniques and applications in industry,

    S. Hayward, K. van Lopik, C. Hinde, and A. A. West, “A survey of indoor location technologies, techniques and applications in industry,” Internet of Things, vol. 20, p. 100608, 2022

    work page 2022

  8. [8]

    Winips: Wifi-based non-intrusive indoor positioning system with online radio map construc- tion and adaptation,

    H. Zou, M. Jin, H. Jiang, L. Xie, and C. J. Spanos, “Winips: Wifi-based non-intrusive indoor positioning system with online radio map construc- tion and adaptation,” IEEE Transactions on Wireless Communications , vol. 16, no. 12, pp. 8118–8130, 2017

    work page 2017

Show all 34 references
  1. [9]

    A high-adaptability indoor localization algorithm for large-scale ble sensors,

    C. Yin, H. Jiang, J. Chen, and X. Miao, “A high-adaptability indoor localization algorithm for large-scale ble sensors,” in 2021 40th Chinese Control Conference (CCC). IEEE, 2021, pp. 5691–5695

    2021

  2. [10]

    Learning progressive distributed com- pression strategies from local channel state information,

    F. Sohrabi, T. Jiang, and W. Yu, “Learning progressive distributed com- pression strategies from local channel state information,” IEEE Journal of Selected Topics in Signal Processing , vol. 16, no. 3, pp. 573–584, 2022

    2022

  3. [11]

    A hybrid gan-inception deep learning approach for enhanced coordinate-based acoustic emission source local- ization,

    X. Huang, M. Han, and Y . Deng, “A hybrid gan-inception deep learning approach for enhanced coordinate-based acoustic emission source local- ization,” Applied Sciences, vol. 14, no. 19, p. 8811, 2024

    2024

  4. [12]

    An efficient normalized rank based svm for room level indoor wifi localization with diverse devices,

    Y . Rezgui, L. Pei, X. Chen, F. Wen, and C. Han, “An efficient normalized rank based svm for room level indoor wifi localization with diverse devices,” Mobile Information Systems, vol. 2017, no. 1, p. 6268797, 2017

    2017

  5. [13]

    A feature-scaling-based k-nearest neighbor algorithm for indoor positioning systems,

    D. Li, B. Zhang, and C. Li, “A feature-scaling-based k-nearest neighbor algorithm for indoor positioning systems,” IEEE Internet of Things Journal, vol. 3, no. 4, pp. 590–597, 2015

    2015

  6. [14]

    Supervised and semi-supervised deep probabilistic models for indoor positioning problems,

    W. Qian, F. Lauri, and F. Gechter, “Supervised and semi-supervised deep probabilistic models for indoor positioning problems,” Neurocomputing, vol. 435, pp. 228–238, 2021

    2021

  7. [15]

    A novel convolutional neural network based indoor localization framework with wifi fingerprinting,

    X. Song, X. Fan, C. Xiang, Q. Ye, L. Liu, Z. Wang, X. He, N. Yang, and G. Fang, “A novel convolutional neural network based indoor localization framework with wifi fingerprinting,” IEEE access , vol. 7, pp. 110 698– 110 709, 2019

    2019

  8. [16]

    Fila: Fine-grained indoor localization,

    K. Wu, J. Xiao, Y . Yi, M. Gao, and L. M. Ni, “Fila: Fine-grained indoor localization,” in 2012 Proceedings IEEE INFOCOM . IEEE, 2012, pp. 2210–2218

    2012

  9. [17]

    Lifs: Low human-effort, device-free localization with fine- grained subcarrier information,

    J. Wang, H. Jiang, J. Xiong, K. Jamieson, X. Chen, D. Fang, and B. Xie, “Lifs: Low human-effort, device-free localization with fine- grained subcarrier information,” in Proceedings of the 22nd Annual International Conference on Mobile Computing and Networking , 2016, pp. 243–256

    2016

  10. [18]

    {ArrayTrack}: A {Fine-Grained} indoor location system,

    J. Xiong and K. Jamieson, “ {ArrayTrack}: A {Fine-Grained} indoor location system,” in 10th USENIX Symposium on Networked Systems Design and Implementation (NSDI 13) , 2013, pp. 71–84

    2013

  11. [19]

    Spotfi: Decimeter level localization using wifi,

    M. Kotaru, K. Joshi, D. Bharadia, and S. Katti, “Spotfi: Decimeter level localization using wifi,” in Proceedings of the 2015 ACM conference on special interest group on data communication , 2015, pp. 269–282

    2015

  12. [20]

    Deepfi: Deep learning for indoor fingerprinting using channel state information,

    X. Wang, L. Gao, S. Mao, and S. Pandey, “Deepfi: Deep learning for indoor fingerprinting using channel state information,” in 2015 IEEE wireless communications and networking conference (WCNC) . IEEE, 2015, pp. 1666–1671

    2015

  13. [21]

    Cifi: Deep convolutional neural networks for indoor localization with 5 ghz wi-fi,

    X. Wang, X. Wang, and S. Mao, “Cifi: Deep convolutional neural networks for indoor localization with 5 ghz wi-fi,” in 2017 IEEE In- ternational Conference on Communications (ICC) . IEEE, 2017, pp. 1–6

    2017

  14. [22]

    Biloc: Bi-modal deep learning for indoor localization with commodity 5ghz wifi,

    X. Wang, L. Gao, and S. Mao, “Biloc: Bi-modal deep learning for indoor localization with commodity 5ghz wifi,” IEEE access, vol. 5, pp. 4209– 4220, 2017

    2017

  15. [23]

    Deep learning based nlos identification with commodity wlan devices,

    J.-S. Choi, W.-H. Lee, J.-H. Lee, J.-H. Lee, and S.-C. Kim, “Deep learning based nlos identification with commodity wlan devices,” IEEE Transactions on Vehicular Technology , vol. 67, no. 4, pp. 3295–3303, 2017

    2017

  16. [24]

    Rf/wifi-based uav surveillance systems: A systematic literature review,

    I. Bisio, C. Garibotto, H. Haleem, F. Lavagetto, and A. Sciarrone, “Rf/wifi-based uav surveillance systems: A systematic literature review,” Internet of Things , p. 101201, 2024

    2024

  17. [25]

    Fusing the wireless technique optical fiber force sensor for remote monitoring of sleeping posture,

    T. Li, Q. Pei, X. Qin, Y . Liang, Y . Tan, and Z. Zhou, “Fusing the wireless technique optical fiber force sensor for remote monitoring of sleeping posture,” IEEE/ASME Transactions on Mechatronics , vol. 28, no. 5, pp. 2703–2715, 2023

    2023

  18. [26]

    Powerskel: A device-free framework using csi signal for human skeleton estimation in power station,

    C. Yin, X. Miao, J. Chen, H. Jiang, J. Yang, Y . Zhou, M. Wu, and Z. Chen, “Powerskel: A device-free framework using csi signal for human skeleton estimation in power station,” IEEE Internet of Things Journal , 2024

    2024

  19. [27]

    Estimating video on demand qoe from network qos through icmp probes,

    G. Miranda, D. F. Macedo, and J. M. Marquez-Barja, “Estimating video on demand qoe from network qos through icmp probes,” IEEE Transactions on Network and Service Management , vol. 19, no. 2, pp. 1890–1902, 2021

    1902

  20. [28]

    Multiclass moisture classification in woodchips using iiot wi-fi and machine learning techniques,

    K. Suthar and Q. P. He, “Multiclass moisture classification in woodchips using iiot wi-fi and machine learning techniques,” Computers & Chemical Engineering, vol. 154, p. 107445, 2021

    2021

  21. [29]

    Comparison between wi-fi-csi and radar-based har,

    A. Dahal, S. Biswas, S. Z. Gurbuz, and A. C. Gurbuz, “Comparison between wi-fi-csi and radar-based har,” in 2024 IEEE Radar Conference (RadarConf24). IEEE, 2024, pp. 1–6

    2024

  22. [30]

    Interactive behavior prediction for heterogeneous traffic participants in the urban road: A graph-neural- network-based multitask learning framework,

    Z. Li, J. Gong, C. Lu, and Y . Yi, “Interactive behavior prediction for heterogeneous traffic participants in the urban road: A graph-neural- network-based multitask learning framework,” IEEE/ASME Transactions on Mechatronics, vol. 26, no. 3, pp. 1339–1349, 2021

    2021

  23. [31]

    Multi-task learning for dense prediction tasks: A survey,

    S. Vandenhende, S. Georgoulis, W. Van Gansbeke, M. Proesmans, D. Dai, and L. Van Gool, “Multi-task learning for dense prediction tasks: A survey,” IEEE transactions on pattern analysis and machine intelligence, vol. 44, no. 7, pp. 3614–3633, 2021

    2021

  24. [32]

    Skeleton-in- context: Unified skeleton sequence modeling with in-context learning,

    X. Wang, Z. Fang, X. Li, X. Li, C. Chen, and M. Liu, “Skeleton-in- context: Unified skeleton sequence modeling with in-context learning,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 2436–2446

    2024

  25. [33]

    A csi-based data-driven localization framework using small-scale training datasets in single-site mimo systems,

    K. Vuckovic, S. Hosseini, F. Hejazi, and N. Rahnavard, “A csi-based data-driven localization framework using small-scale training datasets in single-site mimo systems,” IEEE Transactions on Wireless Communica- tions, 2024

    2024

  26. [34]

    Self-supervised learning for user localization,

    A. Dash, J. Gu, G. Wang, and N. Ansari, “Self-supervised learning for user localization,” arXiv preprint arXiv:2404.15370 , 2024

    2024 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.