The Toronto older adults gait archive: video and 3D inertial motion capture data of older adults’ walking

Sina Mehdizadeh, Hoda Nabavi, Andrea Sabo, Twinkle Arora, Andrea Iaboni, Babak TaatiScientific 2022 · DOI

This research paper, published in Scientific Data, introduces a specialized dataset designed to bridge the gap between high-end laboratory motion capture and accessible, camera-based gait analysis for the elderly.

TL;DR

The Toronto Older Adults Gait Archive provides a rare, open-access dataset of 14 older adults (average age 86.7) featuring synchronized 2D video and 3D inertial motion capture. The study demonstrates that while current computer vision models can accurately track when a senior steps (temporal data), they still struggle to accurately measure how wide or how variable those steps are (spatial data) compared to gold-standard sensors.

What Problem Does This Paper Address?

Falls are a leading cause of injury and mortality in adults over 65. While gait analysis is a powerful predictor of fall risk, the "gold standard" involves expensive 3D motion capture systems or pressure-sensitive walkways that require specialized staff and dedicated lab space.

Computer vision (CV) offers a low-cost alternative using standard RGB cameras. However, most CV pose-estimation models (like OpenPose) are trained on datasets of young, healthy individuals. Older adults often have distinct gait patterns—shorter steps, lower speeds, and different postures—that can cause these algorithms to fail. To fix this, researchers need "ground truth" data that pairs standard video with high-precision motion sensors specifically for the geriatric population.

Methods

The researchers recorded 14 participants walking for 60 seconds in a "front-view" and "back-view" circuit.

  • Hardware: They used two budget-friendly smartphones (Motorola Moto G5) mounted at different heights: 111 cm (eye-level) and 205 cm (ceiling-mounted) to mimic real-world clinical or nursing home environments.
  • Ground Truth: Participants wore an Xsens MVN Awinda system, consisting of seven wireless inertial measurement units (IMUs) attached to the lower body and sacrum.
  • CV Processing: The team processed the videos through three popular pose-estimation libraries: AlphaPose, OpenPose, and Detectron.
  • Clinical Benchmarks: To ensure the data was medically relevant, they also recorded participants' scores on standard clinical tests, including the Berg Balance Scale (BBS) and the Timed Up-and-Go (TUG) test.

Datasets

The authors introduced the Toronto Older Adults Gait Archive, hosted on Figshare. It includes:

  • Raw MP4 video files (with faces blurred for privacy).
  • 3D global coordinates from the Xsens IMU system (CSV and .bvh formats).
  • 2D joint positions (in pixels) extracted by the three CV libraries.
  • Metadata containing demographic info and clinical assessment scores.

This dataset is unique because it focuses specifically on the "old-old" demographic (up to age 99) and provides multiple camera angles for the same walking bouts.

Key Insights and Ideas

  • Timing is Everything: The most significant "aha" moment was that temporal variables (cadence and step time) calculated from video were highly reliable. They correlated strongly with clinical fall-risk scores, matching the performance of the expensive IMU system.
  • The Spatial Struggle: Spatial variables (like step width) and gait variability (how much steps differ from one another) showed poor correlation between the video and the IMU system. This suggests that current CV models aren't yet precise enough to catch the subtle "shuffling" or irregularities that are hallmarks of high fall risk.
  • Perspective Normalization: Since 2D video lacks depth, the researchers normalized spatial measures by the participant's hip width. This is a clever, low-tech way to allow for quantitative comparisons across different camera distances.

Key Results

  • Cadence Accuracy: Cadence (steps per minute) measured via video showed a high correlation with the Timed Up-and-Go (TUG) test ($R \approx -0.8$ to $-0.9$).
  • Consistency: The results were consistent across both the top-mounted and bottom-mounted cameras, suggesting that ceiling-mounted security cameras in nursing homes could potentially be used for gait monitoring.
  • Validation: The study confirmed that while CV can tell you how fast someone is walking, it cannot yet reliably tell you how steady they are (variability) in a frontal plane.

Limitations and Future Work

The primary limitation is the sample size; due to the COVID-19 pandemic, recruitment was cut short at 14 participants (with only 11 used for the final analysis). Additionally, the cameras and IMU systems were not hardware-synchronized (i.e., triggered by the same clock), though they were aligned manually.

The authors suggest that future work should focus on using this dataset to "fine-tune" CV models. By training algorithms specifically on the joint movements of older adults provided in this archive, engineers can develop more robust tools for long-term, automated fall-risk monitoring in home and care settings.

Summary updated

Abstract

We introduce the Toronto Older Adults Gait Archive, a gait dataset of 14 older adults containing 2D video recordings, and 2D (video pose tracking algorithms) and 3D (inertial motion capture) joint locations of the lower body. Participants walked for 60 seconds. We also collected participants' scores on four clinical assessments of gait and balance, namely the Tinneti performance-oriented mobility assessment (POMA-gait and -balance), the Berg balance scale (BBS), and the timed-up-and-go (TUG). Three human pose tracking models (Alphapose, OpenPose, and Detectron) were used to detect body joint positions in 2D video frames and a number of gait parameters were computed using 2D video-based and 3D motion capture data. To show an example usage of our datasets, we performed a correlation analysis between the gait variables and the clinical scores. Our findings revealed that the temporal but not the spatial or variability gait variables from both systems had high correlations to clinical scores. This dataset can be used to evaluate, or to enhance vision-based pose-tracking models to the specifics of older adults' walking.
BibTeX
@inproceedings{mehdizadeh2022toronto,
  author = {Sina Mehdizadeh and Hoda Nabavi and Andrea Sabo and Twinkle Arora and Andrea Iaboni and Babak Taati},
  title = {{The Toronto older adults gait archive: video and 3D inertial motion capture data of older adults’ walking}},
  booktitle = {Scientific},
  year = {2022},
  doi = {10.1038/s41597-022-01495-z},
}
Figure