Concurrent validity of human pose tracking in video for measuring gait parameters in older adults: a preliminary analysis with multiple trackers, viewing angles, and walking directions

Sina Mehdizadeh, Hoda Nabavi, Andrea Sabo, Twinkle Arora, Andrea Iaboni, Babak TaatiJournal of NeuroEngineering and Rehabilitation 2021 · DOI

Abstract

There are important opportunities to evaluate models capable of 3D pose estimation in video data, improve the training of pose-tracking algorithms for older adult and clinical populations, and develop video-based 3D pose trackers specifically optimized for quantitative gait measurement.

Introduction

Gait analysis serves as a critical predictor for fall risk and cognitive decline in older populations. Traditionally, the "gold standard" for such analysis involves expensive 3D infrared motion capture systems or specialized pressure walkways. These systems are often limited to laboratory settings and require expert operation. While depth cameras like the Microsoft Kinect offered a more affordable alternative, their range is restricted to approximately 4.5 meters.

This study explores the clinical validity of using standard RGB video cameras—such as those found in smartphones—combined with deep-learning-based pose trackers. The primary challenge lies in determining if these algorithms, often trained on healthy young adults, can accurately capture the unique postures and slower walking speeds characteristic of older adults.

Methods

The researchers conducted a preliminary study with 11 older adults (average age 85.2 years) walking a 13-meter path. Data was captured using two smartphone cameras positioned at different heights: eye-level (111 cm) and a ceiling-mounted "surveillance" view (205 cm).

Ground truth measurements were obtained using the Xsens MVN Awinda system, which utilizes seven wearable inertial measurement units (IMUs). The video data was processed through three distinct deep-learning pose tracking architectures:

  1. OpenPose: A bottom-up approach detecting all body parts before associating them with individuals.
  2. AlphaPose: A top-down approach that first identifies a person's bounding box and then locates joints within it.
  3. Detectron: A top-down architecture that predicts bounding boxes and keypoints simultaneously.
Comparison of AlphaPose (A), OpenPose (B), and Detectron (C) trackers across front-facing and back-facing walking directions.

The study evaluated six gait variables, including temporal measures (cadence, step time) and spatial measures (step width, estimated Margin of Stability [eMOS]).

Results

The analysis revealed a significant performance gap between temporal and spatial gait parameters.

Temporal vs. Spatial Accuracy

Temporal measures demonstrated near-perfect correlation with the gold standard, with cadence and step time achieving R-values between 0.99 and 1.00 when averaged over a walk. This high accuracy is attributed to the trackers' ability to consistently identify peaks in time-series data (e.g., the lowest point of an ankle during a step), regardless of absolute pixel precision.

In contrast, spatial measures like step width and eMOS were significantly less reliable, with correlations typically ranging from R = 0.30 to 0.60 for individual steps. Estimating 3D coordinates from a 2D plane introduces errors due to perspective and parallax that current off-the-shelf trackers struggle to overcome.

Algorithm and Setup Performance

  • Architecture Matters: Top-down trackers (AlphaPose and Detectron) outperformed the bottom-up OpenPose model, suggesting that isolating the individual first provides a more stable signal for clinical gait analysis.
  • Aggregation Benefits: Correlations improved when data was averaged across an entire walk rather than analyzed step-by-step, following Robinson’s paradox where aggregate data appears more reliable than individual points. For clinicians, this means video tracking is currently better for assessing overall gait quality than for diagnosing single-step anomalies.
  • Camera Versatility: Surprisingly, camera height (eye-level vs. ceiling) had minimal impact on accuracy, suggesting that existing infrastructure like hallway security cameras could potentially be repurposed for gait monitoring.

Discussion and Limitations

While 2D pose trackers are ready for measuring walking speed and cadence in clinical settings, they are not yet reliable for measuring balance and stability metrics like step width. A major limitation of this study was the small sample size (n=11) due to pandemic-related interruptions. Additionally, 2D tracking inherently struggles with occlusion; in back-view videos, the trailing leg often obscures the leading leg, causing the tracker to lose joint positions.

Future work should focus on developing 3D pose estimation models specifically trained on clinical populations to provide the spatial accuracy necessary for effective fall prevention and stability assessment.

Summary updated

Abstract

There are important opportunities to evaluate models capable of 3D pose estimation in video data, improve the training of pose-tracking algorithms for older adult and clinical populations, and develop video-based 3D pose trackers specifically optimized for quantitative gait measurement.
BibTeX
@article{mehdizadeh2021concurrent,
  author = {Sina Mehdizadeh and Hoda Nabavi and Andrea Sabo and Twinkle Arora and Andrea Iaboni and Babak Taati},
  title = {{Concurrent validity of human pose tracking in video for measuring gait parameters in older adults: a preliminary analysis with multiple trackers, viewing angles, and walking directions}},
  journal = {Journal of NeuroEngineering and Rehabilitation},
  year = {2021},
  doi = {10.1186/s12984-021-00933-0},
}
Figure