Automating Timed Up and Go Phase Segmentation and Gait Analysis via the tugturn Markerless 3D Pipeline

This paper introduces \textit{tugturn.py}, a reproducible Python-based pipeline that automates 3D markerless Timed Up and Go analysis by segmenting trial phases, detecting gait events, and calculating comprehensive spatiotemporal and stability metrics to support clinical and research decision-making.

The Gist

Imagine you are trying to understand how a person walks, turns, and sits down. In the past, to get a detailed "biomechanical report card" on their movement, they had to wear a suit covered in reflective stickers (markers) and walk into a high-tech lab filled with expensive cameras. It was like filming a movie in a Hollywood studio: great quality, but expensive, slow, and hard to do outside the studio.

This paper introduces a new tool called tugturn.py (part of a larger system called TUGTURN). Think of this tool as a smart, automated film editor that can watch a regular video of a person doing the "Timed Up and Go" (TUG) test and automatically break it down into a detailed scientific report—without any stickers or special labs.

Here is how it works, using simple analogies:

1. The Problem: The "Blind Stopwatch"

Traditionally, doctors use a stopwatch for the TUG test. They start the timer when the patient stands up and stop it when they sit back down.

• The Flaw: This is like judging a marathon runner only by their total finish time. You don't know how they ran. Did they stumble? Did they turn awkwardly? Did they walk fast but stand up slowly? The stopwatch gives you one number, but it hides the story of the movement.

2. The Solution: The "Smart Traffic Controller"

The tugturn.py software acts like a smart traffic controller for the video. Instead of just watching the whole race, it knows exactly where the "checkpoints" are.

• Spatial Segmentation (The Invisible Fence): The software looks at the video and draws invisible lines on the floor.
  • Line A: The chair area.
  • Line B: The turning area (about 4.5 meters away).
  • Line C: The return path.
  • The Magic: As soon as the person crosses these lines, the software automatically slices the video into five distinct chapters: Stand Up, Walk Forward, Turn Around, Walk Back, and Sit Down. It stops guessing and starts knowing exactly which part of the movement it is analyzing.

3. Finding the Steps: The "Detective"

Once the video is sliced into chapters, the software needs to find exactly when the person's heel hits the ground (Heel Strike) and when they lift their toe (Toe Off).

• The Challenge: In a normal walk, this is easy. But in the TUG test, the person walks forward, turns 180 degrees, and walks back. Old software gets confused during the turn and thinks every wiggle is a step.
• The Fix: The software uses a "Dynamic Detective" approach. It doesn't just look at the foot; it looks at the direction the person is facing. It creates a moving "arrow" that follows the person's path. It only counts a step if the foot moves in the direction of that arrow. If the person is turning or standing still, the detective says, "Not a step!" and ignores it. This prevents false alarms.

4. The Report Card: More Than Just Time

Once the steps are counted, the software generates a rich report card that includes:

• The "Dance" of the Body (Vector Coding): It analyzes how the hips and shoulders move together. If a person with Parkinson's disease turns "all in one block" (like a robot), the software spots this specific pattern. It's like a dance instructor noticing if two dancers are moving in sync or out of sync.
• The "Tipping Point" (XCoM): It calculates a "stability score." Imagine a tightrope walker; this metric predicts if they are about to fall before they actually do. It helps doctors see if a patient is dangerously unstable even if they didn't fall.
• The "Visual Story" (HTML Reports): Instead of giving doctors a boring spreadsheet, it creates a colorful, interactive webpage with graphs and GIFs. It's like turning a dry math homework assignment into a Netflix documentary about the patient's movement.

5. Why This Matters

• No Special Gear: You don't need a $50,000 lab. You just need a camera (like a phone or a webcam).
• Speed: It can process hundreds of videos in minutes, like a batch photo editor, rather than a human watching them one by one.
• Reproducibility: Because it's a computer program, it never gets tired or distracted. It analyzes every patient exactly the same way, making the data reliable for research.

The Bottom Line

This paper presents a digital Swiss Army knife for movement analysis. It takes a simple, everyday test (getting up, walking, turning, and sitting) and uses smart computer vision to turn it into a deep, scientific investigation of how a person moves. It bridges the gap between "just watching a patient" and "understanding the physics of their movement," making advanced biomechanics accessible to any clinic with a camera and a computer.

Technical Summary

1. Problem Statement

The Timed Up and Go (TUG) test is a standard clinical tool for assessing mobility, integrating postural transitions, walking, turning, and sitting. However, current analysis methods face significant limitations:

• Traditional Methods: Rely on specialized optoelectronic motion capture, instrumented walkways, or force plates, which are expensive, infrastructure-heavy, and time-consuming.
• Existing Markerless Pipelines: While computer vision (e.g., MediaPipe) has made 3D pose estimation accessible, existing tools often lack TUG-specific workflows. They typically provide generic gait summaries or pose extraction but fail to:
  • Automatically segment the trial into distinct functional phases (Stand, Gait, Turn, Sit).
  • Accurately detect gait events (heel-strike/toe-off) in bidirectional tasks where turning causes false positives.
  • Provide phase-specific biomechanical metrics (e.g., coordination during turns vs. straight walking).
• Manual Processing: Without automated tools, clinical analysis often resorts to subjective visual inspection or labor-intensive manual frame-by-frame processing, reducing reproducibility and scalability.

2. Methodology: The tugturn.py Pipeline

The authors present tugturn.py, a Python-based, command-line interface (CLI) driven software pipeline designed for markerless 3D TUG analysis. It processes 3D coordinate time-series data (typically derived from MediaPipe via the vailá framework) through a deterministic, multi-stage workflow.

A. Spatial Phase Segmentation (The "Y-Axis Logic")

Instead of treating TUG as a temporal "black box," the pipeline uses spatial thresholds to automatically slice continuous trials into five discrete phases:

1. Sit-to-Stand
2. First Gait (Walking away)
3. Turning
4. Second Gait (Walking back)
5. Stand-to-Sit

• Mechanism: It tracks the estimated Center of Mass (pelvis) along the anteroposterior axis (Y-axis).
• Thresholds:
  • Chair zone: $Y \le 1.125$ m.
  • Turning zone: $Y \approx 4.5$ m.
• This transforms a subjective duration test into a precise topographical examination of movement.

B. Gait Event Detection (Dynamic Zeni Algorithm)

To detect Heel-Strike (HS) and Toe-Off (TO) events without false positives during turns, the pipeline implements a robust adaptation of the Zeni relative-distance algorithm:

1. Walking Mask: Verifies the subject is standing (pelvic Z-height) and moving (XY velocity > 0.15 m/s) before attempting detection.
2. Dynamic Progression Vector: Calculates a smoothed derivative of the pelvic center to generate a dynamic unit directional vector, allowing the algorithm to adapt frame-by-frame to the patient's actual walking direction (forward or return).
3. Event Logic:
   • HS: Projects the vector distance between the heel and pelvis onto the dynamic progression vector; local maxima indicate strikes.
   • TO: Identifies local minima of the toe marker's projection relative to the pelvis.
4. Constraints: Enforces physiological limits (e.g., no consecutive strikes of the same foot within 300ms) and spatial restrictions (discarding events falling outside the "First Gait" or "Second Gait" zones).

C. Advanced Kinematics & Coordination

Beyond basic spatiotemporal metrics, the pipeline extracts:

• Vector Coding: Quantifies intersegmental coordination between the trunk and pelvis during the turning phase. It categorizes behavior into In-Phase (associated with Parkinsonian en bloc turning) or Anti-Phase patterns.
• Dynamic Stability: Calculates Extrapolated Center of Mass (XCoM) metrics to assess stability.
• Trunk Inclination: Monitors trunk angles to detect abnormalities like camptocormia.

3. Key Contributions

• Specialized Workflow: The first widely available, full-body 3D landmark framework custom-built for automated clinical phase segmentation in TUG.
• Robust Event Detection: A novel "Dynamic Zeni" approach that solves the challenge of detecting gait events in bidirectional, turning tasks where standard algorithms fail.
• Reproducibility & Scalability:
  • Configured via TOML files for easy parameter tuning.
  • Supports batch processing of entire clinical databases (hundreds of trials) in minutes.
  • Generates machine-readable outputs (CSV, JSON) and human-readable interactive HTML reports with embedded visualizations.
• Open Source: The software is written in pure Python with minimal dependencies (NumPy, Pandas, SciPy, Plotly), ensuring cross-platform compatibility.

4. Results and Outputs

The paper validates the software using a test dataset (subject s26_m1_t1), demonstrating the following outputs:

• Phase Segmentation: Successfully isolated phases with specific durations (e.g., Turn duration: 2.79s; First Gait: 5.81s).
• Spatiotemporal Metrics: Generated precise data including cadence (67.91 steps/min), velocity (0.347 m/s), and step counts per phase.
• Stability & Coordination: Calculated XCoM deviations (0.027m/0.041m) and Vector Coding metrics.
• Artifact Generation: Produced comprehensive HTML reports, GIF visualizations of segmentation quality, and detailed CSV tables for statistical analysis (e.g., bd_kinematics, bd_vector_coding).

5. Significance and Future Work

• Clinical Impact: tugturn.py bridges the gap between raw video tracking and robust biomechanical endpoints, enabling high-throughput, objective analysis in rehabilitation and uncontrolled clinical settings without expensive hardware.
• Digital Biomarkers: By isolating specific phases (like turning), the tool facilitates the extraction of digital biomarkers (e.g., In-Phase turning) that correlate with disease severity (e.g., Hoehn & Yahr stages in Parkinson's).
• Limitations & Future Directions:
  • Current limitations include dependence on the quality of upstream pose estimation and the need for protocol-specific threshold tuning.
  • Future work aims to include cross-dataset validation, uncertainty quantification, and multicenter reproducibility studies against instrumented gold standards.

In conclusion, tugturn.py represents a significant step toward democratizing advanced biomechanical analysis, making it accessible, reproducible, and scalable for clinical decision-making.