Hippony Cutoff by Hippony LabBiomechanical signal analysis
BIOMECHANICS / SIGNAL QUALITY

Find suitable
cutoff frequencies.

Upload a C3D to compare marker and ground reaction force cutoff frequencies. Get suggested settings, supporting plots and an explanation of their limitations.

Markers only

Get a marker cutoff suggestion from movement data.

Markers + forces

Check marker and force settings together.

Optional joint moments

With supported calibration, compare how filter settings affect joint moments.

  1. 1 Upload
  2. 2 Review
  3. 3 Results
START A RECORDING

Upload your C3D

Gait presets + general diagnostics

Choose a continuous C3D and review its preprocessing, channels and units. Use walking/running presets for gait, or General movement for marker diagnostics of other movements, including landing. General mode does not support force or joint-moment recommendations.

What to prepare before uploading
  • Marker diagnostics: a C3D with physical marker coordinates and stated length units. Point labels may also contain computed outputs: exclude those. Channels with missing frames are excluded from this signal screen; it does not fill gaps.
  • Force checks: calibrated force-platform signals in newtons (N), known global vertical axis and sign, checked synchronization, and at least two complete interior contacts. Upload the continuous recording, rather than separately cut stance clips.
  • Experimental joint moments: the same participant’s static C3D, measured mass (kg) and height (m), and the supported physical marker protocol below. Ordinary marker names or mass and height alone do not provide a calibration.

Public examples include downloaded recordings and, where supported, participant calibration. Review the confirmations yourself: selecting an example does not confirm its validity.

Checking how this analysis server stores uploaded files…

OR TRY A PUBLIC RECORDING

Run a real recording to explore your own settings. For an immediate explanation with saved measurements, choose “Public-data examples” above.

PUBLIC DATA · SAVED EXAMPLES

Public-data examples

Saved Leeds squat · concise report with figures ↗

Public General-mode example: bounded marker diagnostics of a squat. Preprocessing history is unknown; force and joint-moment recommendations are unsupported.

Saved countermovement jump · concise report with figures ↗

Public General-mode example: bilateral countermovement jump, participant 01. Unfiltered trajectories are source-reported; initial labeling and gap filling may be present. Marker diagnostics only; force and joint-moment recommendations are unsupported.

Saved Winter comparison · nine running trials

These public-data examples contain saved measurements. Select a trial to inspect its results; recommendations for your current file appear separately above.

Read the plots · fixed RBDS001runT35 example

All four figures below show participant 001 running at 3.5 m/s. The recording selector changes the numerical evidence above; these figures remain this same example. Read marker filtering, force preservation and calculated moments together.

1 · Marker motion: how much movement does filtering remove?
Saved RBDS001runT35 marker residual RMS versus cutoff with Winter primary median and tail sensitivity
The marker residual is raw position minus filtered position. The left panel shows raw and 21 Hz filtered L.Heel.Bottom global-Y position against recording time in seconds. The right panel summarizes the root-mean-square (RMS) size of removed movement in millimetres versus cutoff frequency in Hz, on a logarithmic vertical scale. Its curve is the median across coordinates and its band is their interquartile spread; the highlighted frequency range shows tail-fit choice sensitivity. The primary Winter estimate is the median of coordinate-specific residual crossings, not a crossing of the displayed median curve. It is 21.47 Hz, rounded to 21 Hz; tail-fit sensitivity is 19.37–22.05 Hz, not a confidence interval. Removed movement can include real motion and measurement noise; this plot does not identify their separate contributions.
2 · External force: compare raw force with 20, 21 and 22 Hz filtering
Saved RBDS001runT35 raw vertical force and filtered force time series at 20, 21 and 22 Hz
The left panel is a force time series around one retained left stance: time in seconds and vertical force in newtons (N). The right panel summarizes worst absolute peak and impulse changes across 82 force contacts at each evaluated cutoff. Compare the measured raw trace with the 20, 21 and 22 Hz traces, especially at contact onset and the peak. At 21 Hz, worst absolute peak and impulse changes over all 82 force-screen contacts are 2.618% and 0.057%; the figure illustrates only its displayed time interval. Impulse is the area under force over time (N·s). Preserving a raw trace does not prove true-force accuracy. The force cutoff is matched to the marker setting as an engineering convention.
3 · Calculated joint moments: the net turning effect at hip, knee and ankle
Actual RBDS001runT35 hip, knee and ankle XYZ moment time series over one left stance at matched 20, 21 and 22 Hz
One left stance at recording time 1.33–1.56 s, using actual native-clock samples and matched 20/20, 21/21 and 22/22 Hz calculations. These are moment time series, not residual curves: time in seconds and net joint moments per body mass in Nm/kg. Showing one stance makes their time course readable; compressing many contacts onto one recording-wide axis can hide this shape. The prototype computes supported stance intervals and masks unsupported intervals, so this is not a complete gait-cycle waveform. A net joint moment is the model’s estimated turning effect from motion and external forces, not an individual muscle force. X, Y and Z are model coordinate components; check the report’s frame definitions before naming anatomical movements.
4 · Average left stance: compare the shape at the same percentage of contact
Average left-stance hip knee ankle XYZ moments across 37 contacts with matched 20 21 and 22 Hz curves and one standard deviation shading at 21 Hz
This view aligns 37 retained left contacts from their first to last retained sample to 0–100% of stance using 101 display points per contact, then shows their mean at each percentage. The horizontal axis is percentage of contact, not seconds or a full gait cycle. Curves compare matched 20/20, 21/21 and 22/22 Hz; shading is ±1 population standard deviation across contacts at 21/21 Hz, not a confidence interval. Both moment views use the same model outputs. Resampling is for this display only; the reported sensitivity metrics use native-clock samples. Differences reflect processing sensitivity and between-contact variation, not known anatomical error.
ACF three-measure checks for this example

The residual-autocorrelation example issued no strict three-measure pair for RBDS001runT35: its marker ranking was unstable and its shortlisted moment comparisons exceeded declared sensitivity limits. Its separate diagnostic calculation used 22/22 Hz. That is a different method and calculation from the saved Winter 21/21 Hz experiment shown above. A provisional starting recommendation can be shown while strict eligibility remains unestablished.

THE BIOMECHANICS

Why a cutoff changes the answer

Methods & equations ↗

1 · Motion

Marker positions estimate segment motion. Velocity and acceleration require differentiation, which amplifies high-frequency fluctuations. Some fluctuations may be noise; some may be real movement.

2 · External forces

Force platforms measure ground reactions. A low cutoff can suppress an impact peak; a high cutoff can retain noise or platform resonance. Preserving a raw peak therefore does not prove the true force was preserved.

3 · Joint moments

Inverse dynamics combines motion, segment mass properties and external forces through force and moment balance. Filtering either input can change the calculated net joint moments.

The practical question: which tested setting survives explicit checks, and which assumptions still need validation? Matching motion and force cutoffs is a tested starting convention here, not a universal physical requirement.

What “cutoff” means in this tool

The pipeline applies an offline, zero-phase Butterworth low-pass filter, order 2 per pass, forward and backward in second-order sections with odd padding. The named cutoff is the final two-pass −3 dB frequency (about 0.707 amplitude), corrected using the exact digital frequency relation. It is not the single-pass design frequency. Forward/backward filtering gives a fourth-order asymptotic magnitude roll-off; it is not a causal fourth-order real-time filter.

Signals are filtered at their native rates before contact summaries. Candidates at or above each signal’s Nyquist frequency (half its sample rate) are excluded. The report records coefficients, padding and comparison supports. Residual ranking trims 1 s at each recording end; signal screens declare their own guards, and moment results use 0.5 s guards at supported segment boundaries. Check the exported supports: the three checks need not retain the same number of contacts.

CURRENT-RUN STRICT CHECKS

How the separate automatic check works

Methods & equations ↗

This strict eligibility check belongs to the current-run ACF workflow. The live Winter estimate is a separate displacement-based starting setting. Its engineering limits can flag a pair for review without establishing that it is inaccurate. Automatic mode considers the first three residual-ranked marker cutoffs, retaining only those with force evidence. A pair uses the same cutoff for motion and force. Mass, height and static calibration are optional for signal screening. Without moment calculation, the conclusion reports the available provisional marker-and-force pair (or a marker-only diagnostic). With moment calculation, it issues a strict three-measure pair only when all required evidence is complete and the rules below pass.

  1. Marker evidence must be stable and interior. Rank raw-minus-filtered residuals by the median across valid marker coordinates of mean squared nonzero-lag autocorrelation (ACF): lower scores are preferred under a noise-like residual assumption. The primary score uses the full interior and a 0.1 s lag window; stability repeats the temporal halves and 0.05, 0.1 and 0.2 s lag windows. The winning frequency must stay unchanged across the tested windows and lag choices and must not lie on the evaluated grid boundary. This is a Challis-inspired adaptation, not an exact replication. The top-three shortlist is an engineering choice, not a confidence interval.
  2. Preserve recorded force features. At least 2 complete interior force contacts are required. Across all retained contacts, worst absolute peak change must be ≤5% and worst absolute impulse change ≤2% relative to unfiltered force by default. Both must pass; you may change these force limits during review.
  3. Check joint-moment sensitivity on common support. Every shortlisted candidate needs a complete assessment on identical fixed, side-assigned contacts: at least 2 complete contacts and at least 3 common matched pairs spanning 5 Hz or more. Against every other compared pair, worst per-contact/joint/component RMS difference must be ≤0.02 Nm/kg and worst absolute positive/negative extrema difference ≤0.05 Nm/kg.

If eligible: choose the highest residual-ranked candidate that passes both force and moment limits. No moment-magnitude or smoothness minimization occurs. If evidence is incomplete, the marker check fails, or no candidate passes both force and moment limits: issue no three-measure pair. A separate diagnostic starting pair may still be available for inspection.

Diagnostic fallback and supported model protocol

The separate marker/force diagnostic prefers the highest residual-ranked matched candidate in the top-three pool passing both force limits. If none passes, it retains the highest-ranked common candidate and reports the failures. Fewer than two force contacts leaves only a marker diagnostic. This fallback cannot become a three-measure recommendation merely because a moment plot was calculated.

Custom uploads can select a supported joint-moment model:

  • Fukuchi physical-marker model: map 30 dynamic roles and all 42 static roles, including the additional anatomical landmarks. Confirm the prescribed pelvis, thigh, shank, heel and foot placements in Fukuchi et al., Table 1 and Figure 1.
  • CAST configurable-cluster model: map 18 static anatomical roles and declare seven tracking groups: pelvis, bilateral thighs, shanks and feet. Each group needs at least three non-collinear measured markers, with every selected tracking marker matched in the static trial. This adapter reports net internal moments in each distal anatomical XYZ frame using its declared joint-centre and inertial assumptions.

Both models require the same participant and marker placement, measured mass and height, and one platform with mm markers and N/Nmm/mm force units in an unchanged positive-global-Y world frame and origin. Check hardware synchronization and every force-to-foot assignment.

Plug-in Gait / CGM1 and CGM2.3 use the pinned pyCGM2 calibration and lower-limb kinetics when its runtime is available. CGM2.3 includes OpenSim inverse kinematics. These workflows require measured CGM anthropometry, a matching static trial, positive global Z up and whole-record force-platform assignments. The fixed upstream 6 Hz COM filter and clinical moment projection remain in the provenance. Renaming markers does not establish equivalence between models.

Marker models, calibration and reporting conventions ↗

Signal-screen selections and contact thresholds do not redefine the moment model’s fixed calibration or contacts. Explicit same-frame rigid-cluster marker reconstruction may be used by the moment workflow; rejected estimates, gaps, exclusions and quality checks remain in its provenance.

The 5%/2% force limits, 0.02/0.05 Nm/kg moment limits, top-three pool and comparison-span requirements are this prototype’s engineering review rules. They are not accuracy limits validated by the cited literature.

How to interpret your conclusion

Conclusion rules — how to read this result
  1. The final provisional three-measure selection ranks the residual-shortlisted matched candidates and requires stable interior marker evidence plus unchanged force and moment-sensitivity limits. Incomplete or failing evidence produces no three-measure pair. A marker/force diagnostic starting pair is reported separately. This heuristic is not a validated moment optimum; it assumes noise-like residuals and matched bandwidth as a starting convention.
  2. A force shortlist needs at least 2 complete interior contacts. Relative to the unfiltered recording, across all retained contacts, worst absolute peak change must be ≤ the entered peak limit AND worst absolute impulse change must be ≤ the entered impulse limit (defaults: 5% and 2%, adjustable). This describes recorded features, not force accuracy.
  3. Marker residuals and smoothness alone cannot choose an optimal cutoff. Tested cutoffs stay below native Nyquist, with the declared edge guard applied.
  4. With a supported participant calibration, the app calculates moments across the declared candidate set. Marker evidence, force preservation and fixed-contact moment sensitivity jointly determine whether a provisional pair is eligible. If evidence is unstable, incomplete or fails the declared limits, no three-measure pair is issued. Reconstructed markers and excluded contacts are reported.
  5. A final pair needs explicit moment comparisons and justified model, noise, and outcome evidence; stable curves alone are insufficient. This app identifies no validated optimum. Your confirmations are workflow assertions, not independent validation.
Full conclusion and analysis limits

No recording-specific conclusion yet. Default force limits are 5% peak and 2% impulse; these can be changed during review.

Scientific context and publication use

Challis (1999) motivates residual-autocorrelation reasoning. Mai & Willwacher (2019), Bisseling & Hof (2006) and Kristianslund et al. (2012) discuss filtering and inverse-dynamics sensitivity. These primary papers support investigating the problem; this tool does not reproduce their methods exactly or validate its own thresholds through them.

What to report in a study

Help readers understand what data you used, which settings you chose, and why. You do not need to explain every calculation in the main text.

  1. Describe your recording. State the movement, participants and equipment. Explain where the markers were placed and how the force measurements were checked. For joint moments, also describe the body model and participant measurements used.
  2. Give both cutoff frequencies. Report the marker setting and the force setting separately, in Hz. Explain whether you chose them yourself or used the app’s provisional suggestion. A lower cutoff removes more rapid changes, which can include both noise and real movement.
  3. Explain why you chose them. Summarize the marker analysis, changes in force peaks and impulse (force accumulated over a contact), and changes in joint moments when you tried other settings. Include the range of plausible settings and any warnings.
  4. Say how much usable data remained. Give the number of contacts used. Mention missing markers, estimated marker positions and any sections left out of the analysis.
  5. Save the supporting results. Keep the report, figures, data tables and settings files. These let another researcher check your reasoning and repeat the analysis.
Keep these two claims separate.

“The result changes little when I change the cutoff” describes sensitivity. “The result is close to the true value” describes accuracy and needs a suitable independent reference. A smooth curve alone does not show that a joint moment is accurate.

Technical details for a methods section or appendix

Include these with the exported files so the analysis can be reproduced:

  • Measurements and model: dataset, participant calibration, marker-to-model assignments, units, coordinate axes, sampling rates and synchronization.
  • Filtering: filter type and order, forward/backward processing, the definition of the reported cutoff, and all frequencies tested. This app uses a Butterworth filter of order 2 per pass, applied forward and backward; reported cutoffs refer to the final two-pass −3 dB response.
  • Data handling: how contacts were detected and assigned to each foot, how much data was excluded near recording or segment boundaries, and how missing markers were reconstructed or excluded.
  • Decision and limitations: the comparison limits used, how the final pair was chosen, model assumptions, and any independent validation. Identify a saved demonstration separately from your own recording’s results.
  • Files to retain: the report and figures, CSV data tables, decision JSON (the machine-readable settings and results), and moment provenance (the record of inputs and calculation steps).