Pith. sign in

REVIEW 3 major objections 6 minor 12 references

Pedestrian Dead-Reckoning Algorithms For Dual Foot-Mounted Inertial Sensors

T0 review · 3 major / 6 minor · reviewed 2026-08-14 · deepseek-v4-flash

Pith's one-line read The paper proposes pedestrian dead-reckoning algorithms that reconstruct closed indoor walking paths from foot-mounted inertial sensors, using either one IMU or two coupled IMUs, and reports that the dual-IMU version reduces the mismatch…

desk verdict Clever dual-IMU ZUPT fusion, but the DTW-based validation partly measures the algorithm's own coupling rather than absolute error; the error-reduction claim needs ground truth. read the letter →

arxiv 1908.07369 v1 pith:T3OZB33I submitted 2019-08-20 eess.SY cs.SYeess.SP

classification eess.SYcs.SYeess.SP
keywords pedestriandead-reckoningfoot-mountedIMUZUPTKalmanfilterRTSsmoothingdynamictimewarpingindoornavigationdualinertialsensors
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

Indoor pedestrian trajectories can be reconstructed from foot-mounted inertial sensors alone, without Wi-Fi, beacons, or GPS, provided the path is closed. The paper's single-IMU algorithm augments a zero-velocity-update Kalman filter with the constraint that start and end positions coincide, and applies RTS smoothing together with angle corrections to remove step-boundary discontinuities. Its dual-IMU algorithm couples the two legs by treating the middle of a step as a moment when the moving foot lies approximately at the position of the stationary foot, feeding that as a pseudo-observation into both filters. In experiments of one to fifteen minutes, the dual-IMU method lowered the DTW distance between the left- and right-foot reconstructions to between roughly 0.25 m and 1.55 m, and the authors conclude the reconstructed trajectories reproduce room geometry well enough to serve as reference paths when no survey-grade ground truth is available.

What carries the argument

The load-bearing mechanism is an error-state Kalman filter with zero-velocity updates: the detector T classifies each foot as stationary, and the filter treats measured velocity during stance as a zero pseudo-observation. A closed-loop variant adds position observations at the initial and final standstill intervals, with an RTS smoother spreading that correction across the whole trajectory. The dual-IMU extension adds a coupling pseudo-observation at the middle of each step, using the current position of the other leg (observation matrix that selects the first two position coordinates) as the measurement. DTW with an index-window constraint w is the comparison tool that aligns, scores, and averages the two leg trajectories.

What would settle it

A direct experiment would be to walk a closed path with surveyed waypoints while a motion-capture system records true foot positions; if the reconstructed trajectory deviates from the surveyed path by more than the reported DTW values, or if the errors concentrate at moments when the two feet are far apart mid-step, then the mid-step equality assumption is violated and the central claim is weakened.

Watch

Extended reading notes

Core claim

The central claim, stated in Section V, is that the proposed algorithms reconstruct real trajectories with significantly reduced error compared with a trajectory from a single inertial system. The discovery is that a closed-loop constraint plus RTS smoothing and angle correction already reduces the DTW mismatch between the two legs' independently computed paths, and that coupling a second foot-mounted IMU with a mid-step position pseudo-observation reduces it further: on the longest tested walk the mismatch fell from 17.917 m with the baseline algorithm to 1.546 m with the dual-IMU algorithm. The paper also shows that DTW with an index-window limit is a more reliable curve-comparison metric than discrete Fréchet distance for this setting, and that averaging the two matched leg trajectories yields a single generalized path.

Load-bearing premise

The dual-IMU algorithm couples the two legs by assuming that at the middle of a step the moving foot is approximately at the same position as the stationary foot, and if that fails, for example on wide strides, turns, stairs, or running, the coupling injects biased position corrections into both legs.

Editorial extensions

If this is right

  • On closed indoor paths, a single foot-mounted IMU with loop-closure and smoothing reconstructs paths whose left–right DTW mismatch is below about 3 m in walks of several minutes.
  • Coupling a second foot-mounted IMU at mid-step further reduces the mismatch, reaching values from 0.252 m to 1.546 m across the reported experiments.
  • The reconstructed paths preserve room dimensions, so they can be used as reference trajectories in indoor environments where no more accurate positioning is available.
  • Restricting DTW point matches to an index window reduces the matching complexity from quadratic time to linear time when the window is fixed, and avoids nonsensical long-range matches.

Reading between the lines

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

  • Because the reported metric compares the left-foot reconstruction with the right-foot reconstruction, it mainly captures relative error; common errors such as a shared heading bias or a uniform scale factor could remain invisible, so validation against surveyed waypoints would be a natural next step.
  • The mid-step foot-equality assumption is most plausible for level walking at a moderate pace; on stairs, running, or turning in place the coupling would inject biased corrections, and a testable extension would gate the pseudo-observation by step length or gait mode.
  • The same coupling idea could be transferred to other wearable sensor pairs, such as shin and waist, but the geometric relationship at the pseudo-observation time would need to be re-derived for each mounting.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes two algorithms for pedestrian dead reckoning from foot-mounted IMUs: a single-IMU Kalman filter with zero-velocity updates (ZUPT) and RTS smoothing that includes an in-loop angle correction, and a dual-IMU algorithm that couples two such filters through a mid-step pseudo-observation of the other foot's position. The dual-IMU algorithm also produces a generalized trajectory using DTW-based averaging. The authors evaluate the algorithms on six indoor walking experiments and report DTW distances between the left- and right-leg reconstructions as the accuracy metric.

Significance. If the claimed accuracy improvement were properly established, the dual-IMU algorithm would be a useful contribution to indoor pedestrian navigation. The paper contributes clear pseudocode for closed-loop ZUPT-aided smoothing, a dual-filter coupling mechanism, and a DTW-based trajectory averaging method with a window constraint that reduces computational complexity. The algorithms are described in sufficient detail to allow reimplementation, and a data link is provided. However, the current validation does not measure error against any known reference path, so the central quantitative claim that the algorithms 'reconstruct real trajectories with significantly reduced error' is not supported by the presented evidence.

major comments (3)
  1. [Section V, Table I] The DTW distance reported in Table I is computed between the left- and right-leg reconstructed trajectories, not between the reconstructed and a known true path. Two trajectories can both deviate from the true trajectory by a common bias (for example, a heading error) and still have a small DTW distance. Since the paper itself states in Section V that 'the values of DTW metric do not give a direct answer to the question about the metric difference of the reconstructed and real trajectories,' the conclusion that the algorithms 'reconstruct real trajectories with significantly reduced error' is not supported by the presented evidence. A ground-truth reference, such as a surveyed path or motion-capture data, is needed to substantiate this claim.
  2. [Section IV, Algorithm 4] The dual-IMU algorithm directly minimizes the metric used for validation. At mid-step, the correction step with H=[I2x2 O2x7] injects the other foot's horizontal position as a pseudo-observation, actively pulling the two reconstructed trajectories together. Therefore the large DTW reductions in Table I (for example, experiment 6 from 17.917 to 1.546) are to a significant extent a by-construction consequence of the coupling, not independent evidence of improved absolute accuracy. This circularity must be addressed, for example by evaluating against an external reference rather than a metric that the algorithm is designed to reduce.
  3. [Section IV, step 3 and Algorithm 4] The mid-step equality assumption—'in the middle of a step the position of a moving foot is supposed to be approximately equal to the position of the other foot'—is physically fragile during turns, stair climbing, running, or unusually long strides. If this assumption is violated, the pseudo-observation injects biased position information into both filters, potentially degrading accuracy. The presented experiments cover only horizontal straight-line and L-shaped corridor walking with durations up to 10.5 minutes, so the range of validity of the dual-IMU algorithm is not demonstrated.
minor comments (6)
  1. [Section V, Table I] There is an inconsistent use of decimal separators in Table I (for example, '93,5' versus '216.7'); the table should use a single convention.
  2. [Section II.A] Equation (5) contains a Cyrillic 'С' in the matrix expression; this appears to be a typo and should be replaced with the appropriate Latin letter.
  3. [Algorithm 4] The function 'count_step_length' is used to determine the mid-step point but is never defined; its definition is important because the pseudo-observation is applied at step_len/2.
  4. [Section III and IV] The values of the tuning parameters (gamma, N, Q, R, R', R'', R''', and the DTW window w) are not provided, so the experiments cannot be reproduced from the text alone; please include them as a table or supplementary material.
  5. [Section III, Figure 2] The caption refers to red and yellow paths, but the figure (as rendered) does not have a legend; please make the color coding explicit in the caption.
  6. [Section IV, step 2] The assumption that the leg making the first step is identifiable because the first step is shorter than the other leg's first step is not validated; if this assumption fails, the initial leg assignment in Algorithm 4 may be wrong.

Circularity Check

2 steps flagged · score 6.0 of 10

The main validation metric (inter-leg DTW) is directly minimized by Algorithm 4's mid-step coupling and by the yaw-alignment fitting, so the reported error reductions are partly by construction rather than independent checks.

  1. self definitional [Section IV, Algorithm 4 (mid-step pseudo-observation) and Section V, Table I]
    "In the middle of a step the position of a moving foot is supposed to be approximately equal to a position of the otherot foot. According to that idea, the additional observations may be passed to the Kalman filter."

    Algorithm 4 implements this idea as a Kalman correction at mid-step (H=[I2×2 O2×7], with 'pos' being the current horizontal position of the other leg), so the two reconstructed trajectories are actively pulled toward each other. The only quantitative validation in Table I is the DTW distance between the reconstructed left- and right-leg trajectories. Therefore the reported improvement of Algorithm 4 (e.g., 17.917 to 1.546 in experiment 6) is the expected consequence of the coupling mechanism, not independent evidence that the trajectories are closer to the real path.

  2. fitted input called prediction [Section IV, 'Algorithm for two IMUs', steps 1 and 5; Section V, Table I]
    "For that purpose, the DTW distance is minimized by selecting the initial angles."

    The same DTW distance that later serves as the validation metric is the objective for fitting the relative yaw alignment of the two reconstructed trajectories; step 5 repeats this by 'optimally overlapped once again using brute-force angle search'. The reported DTW values are thus minimized over the alignment parameters before being presented as reconstruction-error evidence. Moreover, DTW between two reconstructed legs, even after optimal overlap, is not a ground-truth error measure: a common heading or scale bias would leave it small while both paths deviate from reality.

full rationale

The paper contains no self-citation chain or imported uniqueness theorem; the algorithms (ZUPT Kalman filter, RTS smoothing, closure post-processing) are standard and described in pseudocode. The circularity lies entirely in the validation logic. Section V concedes that 'DTW metric do not give a direct answer to the question about the metric difference of the reconstructed and real trajectories', yet immediately uses the DTW reductions in Table I to state that the algorithms 'reconstruct real trajectories with significantly reduced error compared to trajectories reconstructed according to a single INS'. For Algorithm 4 the problem is sharper: the mid-step pseudo-observation is expressly designed to make the two legs' positions coincide, so the inter-leg DTW distance is reduced by construction. In addition, the initial-yaw overlap needed to compare the legs is obtained by minimizing the same DTW quantity, so the reported numbers are fitted values rather than independent measurements. There is no ground-truth reference (markers, surveyed path, or external positioning) in the experiments, and the qualitative remark that trajectories 'reproduce the metric characteristics of rooms' is not a quantitative accuracy check. Because the central claim of reduced real-trajectory error rests on a metric that Algorithm 4 itself minimizes, the score is 6: partial circularity, with the single-IMU Algorithm 3 retaining some independent algorithmic content (closure smoothing) but still being validated only by inter-leg DTW rather than absolute error.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

The paper introduces no new physical entities. The central algorithms rely on ZUPT, closed-loop endpoints, and the mid-step foot-position equality, plus several unspecified Kalman tuning parameters. None of the numerical values for these parameters are reported, which is the main reproducibility gap.

free parameters (6)
  • Zero-velocity detection threshold gamma = not reported
    In eq. (9), stationary foot detection uses T < gamma; the value is never specified, yet it controls when velocity pseudo-observations are applied.
  • Time window length N for stationary detection = not reported
    The zero-velocity detector uses a window W_n of N samples centered at t_n; N is not given.
  • Process noise covariance Q = not reported
    The Kalman filter requires Q (6x6 measurement noise), but its value is not stated.
  • Observation noise covariances R, R', R'', R''' = not reported
    The filter uses different R matrices for stationary velocity, standstill position+velocity, and cross-leg position; values are not given.
  • Mid-step fraction (step_len/2) = 0.5 of a step
    Algorithm 4 applies the cross-leg position correction exactly halfway through a step; this hand-chosen timing is not justified.
  • DTW window constraint w = not reported
    The maximum index difference w is introduced to improve matching but its exact value is left to judgment, described only as near the average step length in samples.
assumptions (6)
  • standard math Kalman filter linearization and RTS smoothing recursions are valid for the nonlinear system (Equations (1)-(8)).
    The paper builds on standard Kalman filtering with a linearized transition matrix F_n; no proof of convergence or observability is given.
  • domain assumption The velocity of a stationary foot is exactly zero (ZUPT), used as a pseudo-observation.
    This is the core assumption of ZUPT-aided INS; it is standard but can introduce bias on uneven terrain or during shuffling.
  • domain assumption The trajectory is closed: the start and end positions coincide (Introduction: 'only closed trajectories are considered').
    Algorithm 3 uses the final standstill position as a correction for the whole trajectory; the method is inapplicable to open paths.
  • domain assumption At mid-step, the moving foot's position is approximately equal to the other foot's position (Section IV).
    This is the fusion premise for Algorithm 4; it is approximate and unquantified, and only two horizontal coordinates are used (H=[I2x2 O2x7]).
  • ad hoc to paper The first step is made from the initial standstill position and is shorter than the first step of the other leg (Section IV, step 2).
    This heuristic is used to identify which leg leads; no data support is provided.
  • standard math DTW and Frechet distance are meaningful proxies for trajectory similarity and can be aligned by brute-force yaw-angle search.
    These are established algorithms, but their use as a validation metric is what the paper assumes.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Pedestrian Dead-Reckoning Algorithms For Dual Foot-Mounted Inertial Sensors." pith.science (2026). https://pith.science/paper/T3OZB33I

@misc{pith2026190807369,
  author       = {Pith},
  title        = {Pith review of: Pedestrian Dead-Reckoning Algorithms For Dual Foot-Mounted Inertial Sensors},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/T3OZB33I}},
  note         = {Machine review of arXiv:1908.07369}
}
read the original abstract

This work proposes algorithms for reconstruction of closed-loop pedestrian trajectories based on two foot-mounted inertial measurement units (IMU). The first proposed algorithm allows calculation of a trajectory using measurements from only one IMU. The second algorithm uses data from both foot-mounted IMUs simultaneously. Both algorithms are based on the Kalman filter and the assumption that while a foot is on the ground its velocity is supposed to be zero. Two methods for comparing the obtained trajectories are proposed, advantages and disadvantages of each method are indicated and a way to optimize the computation time is presented. In addition, a method is proposed for constructing one generalized trajectory of human motion based on the trajectories of each leg.

Figures

Figures reproduced from arXiv: 1908.07369 by the authors.

Figure 1
Figure 1. Placement of measurement units. Both the position and the orientation of the devices can also be corrected if there are data from two different IMUs fixed respectively on the right and left legs. The advantage of using multiple sensors has been shown, for example, in reference [7]. In this work, we use the additional as￾sumption that positions of the sensors in space cannot significantly differ from each other. An a… view at source ↗
Figure 2
Figure 2. Comparison of trajectories with preliminary correction of [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Comparison of the metrics Incorrect mapping takes place when turning points of trajectories are at a significant distance from each other. In some cases points that are far apart in time can be mapped by an algorithm although such mapping is undesirable. Thus, the results of the algorithms can be improved by using an additional parameter w — the maximum allowed difference between the indices of the corresponding poi… view at source ↗
Figures from the paper (3 more)
Figure 5
Figure 5. Figure 5: Comparison of metrics with the use of additional [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Comparison of algorithms [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Comparison of algorithms. VI. Conclusion The paper contains description of the algorithms for reconstruction of close-loop trajectories based on infor￾mation about acceleration and angular velocity (for only one IMU and for IMUs on both legs). The corresponding pseudoc…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

12 extracted references · 12 canonical work pages

  1. [1]

    S., Nilsson, J.-O., and H\"andel, P., Smoothing for ZUPT-aided INSs, 2012 International Conference on Indoor Positioning and Indoor Navigation (IPIN), Sydney, IEEE, 2012, pp

    Colomar, D. S., Nilsson, J.-O., and H\"andel, P., Smoothing for ZUPT-aided INSs, 2012 International Conference on Indoor Positioning and Indoor Navigation (IPIN), Sydney, IEEE, 2012, pp. 1-5

  2. [2]

    , Skog, I., H\"andel, P., and Hari, K

    Nilsson, J.-O. , Skog, I., H\"andel, P., and Hari, K. V. S., Foot-mounted INS for everybody --- an open-source embedded implementation, Proceedings of the 2012 IEEE/ION Position, Location and Navigation Symposium, Myrtle Beach, IEEE, 2012, pp. 140-145

  3. [3]

    K., and H\"andel, P., Foot-mounted inertial navigation made easy, 2014 International Conference on Indoor Positioning and Indoor Navigation (IPIN), Busan, IEEE, 2014, pp

    Nilsson, J.-O., Gupta, A. K., and H\"andel, P., Foot-mounted inertial navigation made easy, 2014 International Conference on Indoor Positioning and Indoor Navigation (IPIN), Busan, IEEE, 2014, pp. 24-29

  4. [4]

    M., Error Analysis of ZUPT-Aided Pedestrian Inertial Navigation, 2018 International Conference on Indoor Positioning and Indoor Navigation (IPIN), Nantes, IEEE, 2018

    Wang, Y., Chernyshoff, A., and Shkel, A. M., Error Analysis of ZUPT-Aided Pedestrian Inertial Navigation, 2018 International Conference on Indoor Positioning and Indoor Navigation (IPIN), Nantes, IEEE, 2018

  5. [5]

    Tjhai, C., and Okeefe, K., Comparing Heading Estimates from Multiple Wearable Inertial and Magnetic Sensors Mounted on Lower Limbs, 2018 International Conference on Indoor Positioning and Indoor Navigation (IPIN), Nantes, IEEE, 2018

  6. [6]

    MPU-9250 Product Specification, Revision 1.1, \\ https://www.invensense.com/wp-content/uploads/2015/02/PS-MPU-9250A-01-v1.1.pdf

  7. [7]

    G., and Galov, A

    Mikov, A. G., and Galov, A. S., Data processing algorithms for MEMS based multi-component inertial measurement unit for indoor navigation, 2017 24th Saint Petersburg International Conference on Integrated Navigation Systems (ICINS), St. Petersburg, IEEE, 2017, pp. 1-7

  8. [8]

    Eiter, T., and Mannila, H., Computing discrete Fr\'echet distance, Technical Report CD-TR 94/64, Christian Doppler Laboratory for Expert Systems, Vienna, 1994

Show all 12 references
  1. [9]

    M\"uller, M., ``Dynamic Time Warping'' in Information Retrieval for Music and Motion, Springer, 2007, pp. 69-84

  2. [10]

    Vaughana, N., and Gabrysa, B., Comparing and combining time series trajectories using Dynamic Time Warping, Procedia Computer Science, 2016, vol. 96, pp. 465-474

  3. [11]

    IEEEtran_HOWTO

    11em plus .33em minus .07em @technote 4000 4000 100 4000 4000 500 `\.=1000 = #1 #1 #1 0pt [0pt][0pt] #1 * \| ** #1 \@IEEEauthorblockNstyle \@IEEEauthorblockAstyle \@IEEEauthordefaulttextstyle \@IEEEauthorblockconfadjspace -0.25em \@IEEEauthorblockNtopspace 0.0ex \@IEEEauthorbl...

  4. [12]

    " z * 2 : N V j r 0210 0100 Autod bZ;624701624701624701 n dd 2019:03:11 16:04:12 2019:03:11 16:04:12d ' R98 0100 ( H H JFIF C C

    11em plus .33em minus .07em 4000 4000 100 4000 4000 500 `\.=1000 = #1 \@IEEEnotcompsoconly \@IEEEcompsoconly #1 * [1] 0pt [0pt][0pt] #1 * [1] 0pt [0pt][0pt] #1 * \| ** #1 \@IEEEauthorblockNstyle \@IEEEcompsocnotconfonly \@IEEEauthorblockAstyle \@IEEEcompsocnotconfonly \@IEEEco...

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.