Pose Analysis Framework

Pose Analysis Quick Guide

This quick guide walks through the main stages of pose analysis, from acquisition, feature selection, and pre-processing, to linear and recurrence analysis. It is designed to help readers identify common options at each stage and choose an approach that fits their data and research question.

Two stylized pose skeletons in different poses
Stage 1

Acquisition

Start by asking what kind of movement you want to capture and whether the recording supports stable temporal analysis.

Choose Source

Marker-based or markerless?

Both can feed the same pipeline as long as the output is frame-wise landmark trajectories.

  • Marker-based: higher spatial precision, more setup burden.
  • Markerless video: more scalable and naturalistic, but noisier and more prone to occlusion.
Geometry

2D single camera or 3D reconstruction?

Most researchers will have access to 2D video data but may want to consider 3D reconstruction.

  • 2D: simpler and lighter, but sensitive to viewpoint and orientation.
  • 3D: reduces perspective distortion, but adds reconstruction noise and dimensionality.
Sampling

Is the frame rate constant?

Note that recurrence analysis assumes a fixed temporal unit for delay embedding.

  • Yes: proceed directly to feature selection.
  • No: store timestamps and plan to resample during preprocessing.
Stage 2

Feature Selection

This step involves choosing a movement time series that offers a good balance of anatomical specificity, minimal redundancy, and interpretability.

Option A

Raw coordinates

Keep (x, y) or (x, y, z) trajectories when specific landmarks are the unit of interest.

  • Best for localized movement questions.
  • Preserves anatomical specificity.
  • Carries more redundancy across nearby landmarks.
Option B

Magnitude vectors

Collapse frame-to-frame movement into orientation-agnostic displacement magnitude.

  • Useful when direction matters less than amplitude.
  • Reduces axis redundancy.
  • Sacrifices directional information.
Option C

Derived kinematic features

Create theoretically guided signals from landmark pairs or groups.

  • Examples: blink aperture, lip opening, joint distances, head rotation.
  • Most interpretable when guided by domain knowledge.
  • Selective by design, so they only sample part of the movement space.
Option D

Regional pooling

Group landmarks into broader regions.

  • Examples: pooling across the face, arms, hands, or whole-body subsets.
  • Useful when a regional summary is more meaningful than individual landmarks.
  • Reduces anatomical precision and can hide meaningful joint-level differences.
Stage 3

Pre-processing

The next step is to clean and standardize the pose data so it is ready for downstream analysis.

Missing Data

How much interpolation is safe?

For recurrence analyses, interpolate short gaps only.

  • Preferred rule: keep gap length below Lmax = (m - 1)τ.
  • Short gaps can be linearly interpolated.
  • Long gaps be excluded at on a window-by-window basis.
Alignment

Do camera geometry and pose orientation vary?

If yes, align trajectories into a shared coordinate system before analysis.

  • Use translation, rotation, and optional scaling via Procrustes.
  • Template can be global, reference-frame based, or synthetic.
  • Visually inspect raw versus aligned poses to verify the transform.
Signal Cleaning

Do you need filtering or detrending?

Only apply these when they remove artifacts more than they distort the temporal pattern.

  • Low-pass filtering reduces jitter from pose tracking noise.
  • Detrending helps when drift is clearly nuisance variance.
Normalization

Is amplitude meaningful?

Choose scaling based on whether you want to preserve movement size or focus on structure.

  • Z-score: best when amplitude is nuisance variance.
  • Unit interval: keeps within-signal proportional structure.
  • Centering only: keep scale when absolute magnitude is important.
Windowing

Should the data be segmented over time?

Windowing helps when movement changes over the course of a recording and a single summary would hide that structure.

  • Useful for longer, nonstationary recordings.
  • Lets downstream analyses track temporal change rather than collapse across a whole trial.
Analysis Selection

Choose an Analysis

Now decide what kind of question you want the analysis to answer.

Do you want to know how much movement occurred, how movement is coordinated in space, or how movement unfolds over time?
Linear metrics can quantify movement amount and broad spatial covariation, whereas recurrence analysis is best for understanding how movement patterns and coordination unfold over time.
Stage 4A

Linear Metrics

Linear analyses quantify movement magnitude and dominant movement modes. They are useful as a first pass and as context for interpreting recurrence findings.

Individual

Amplitude-based movement summaries

Use direct kinematic summaries when the question is about how much movement happened.

  • RMS displacement
  • Mean velocity and acceleration
  • Standard deviation of velocity or acceleration
  • Velocity or acceleration magnitude for direction-free summaries
Dyadic

Cross-correlation based movement summaries

Use paired linear measures when the question is about how two people covary over time.

  • Cross-correlation or lagged correlation between matched individual signals.
  • Useful for asking whether partners speed up, slow down, or fluctuate together.
  • Best interpreted alongside the individual movement summaries.
Structure

Principal Components Analysis

Use PCA to reveal principal movement patterns and to diagnose preprocessing quality.

  • Great for spatial covariance and dominant postural modes.
  • Can be computed at the individual or group level.
Best when Comparing movement amount, speed, or variability across conditions.
Strengths Interpretable and easy to link to task manipulation.
Limitations Two signals can share the same amplitude statistics but differ in temporal organization.
Stage 4B

Recurrence Analysis

Recurrence methods answer the temporal question: when does a system return to similar states, and how are those returns organized over time?

Type

Choose the recurrence type

Select the mode that matches the unit of analysis.

  • Auto-RQA: one signal or one participant over time.
  • Cross-RQA: two signals or two interactants in shared dynamics.
  • Multidimensional RQA: multiple signals treated as one system.
Embedding

Set state-space parameters

Delay embedding reconstructs the trajectory before recurrence is computed.

  • Delay τ: estimate with AMI and test plausible ranges rather than relying on a single value.
  • Dimension m: estimate with FNN; movement studies often converge around 3 to 5.
  • Check robustness across a small parameter grid.
Detection

Define what counts as recurrence

Determines the density and interpretability of the recurrence plot.

  • Radius ε: fixed threshold or tuned to target RR, often around 2 to 5% recurrence.
  • Rescaling: mean-rescaled distances help compare across heterogeneous signals.
  • Theiler window: mainly for auto-RQA to suppress trivial temporal neighbors.
  • Minimum line length: increase this value to avoid artificially inflated determinism.
Outputs

Interpret the metrics

RQA returns metrics that capture different aspects of the recurrence structure. Commonly used metrics include:

  • RR / %REC: how often states recur.
  • DET: how predictable the repeated sequences are.
  • Lmax: the longest diagonal line, reflecting the most sustained repeated sequence.
  • Entropy: complexity of line-length structure.
Best when Investigating temporal organization, stability, coupling, or repeated movement patterns over time.
Strengths Captures nonlinear and nonstationary structure that amplitude-based summaries can miss.
Limitations Sensitive to choices about embedding, thresholds, and windowing, so parameter checks are important.
This quick guide is designed to support the following paper: Sale, C., Macpherson, M. C., Patil, G., Miles, K., Kallen, R. W., Wallot, S., & Richardson, M. J. (2026). Nonlinear methods for analyzing pose in behavioral research. arXiv:2604.01453.