TL;DR
This paper investigates using everyday video cameras and advanced pose-estimation software to measure gait in individuals with Parkinson's Disease (PD), comparing it to a clinical gold-standard instrumented walkway. The findings suggest that 2D pose-estimation libraries can capture gait trends reliably, offering a promising avenue for remote, longitudinal monitoring of PD progression outside of clinical settings.
What Problem Does This Paper Address?
Parkinson's Disease (PD) significantly impacts gait, leading to changes like slower walking, shorter steps, and increased variability. Monitoring these changes is crucial for tracking disease progression and the effectiveness of treatments. While clinical settings use sophisticated systems like instrumented walkways, these are expensive and inaccessible for frequent, at-home monitoring. This creates a need for more accessible, vision-based solutions that can capture gait data reliably in non-clinical environments. Previous research has explored body-worn sensors and even basic video analysis, but validating these against clinical standards for PD patients, especially using modern pose-estimation techniques, remains an active area of research.
Methods
The study employed three popular 2D human pose-estimation libraries (AlphaPose, Detectron, and OpenPose) to extract joint trajectories from color videos of participants walking. They also explored a 3D pose-estimation library (ROMP) for comparison. A critical step was identifying "heel-strike" events, which mark the beginning of a step. This was done using both an automated clustering algorithm (ST-DBSCAN) and manual annotation by researchers.
Gait features were then calculated from both the video data and a Zeno instrumented walkway (considered a clinical standard). These features included:
- Number of steps
- Cadence
- Mean and coefficient of variation (CV) of step width
- Mean and CV of step time
- Gait speed and step length (only for 3D ROMP)
The core of the study involved a correlation analysis to assess how well the video-derived gait features aligned with those from the Zeno walkway. They specifically analyzed walking towards and away from the camera separately, as well as comparing automated vs. manual heel-strike detection.
Datasets
The study utilized data collected from 25 adults diagnosed with idiopathic Parkinson's Disease. Participants walked 6 meters along a Zeno instrumented walkway, turned, and walked back, all while being recorded by a standard color camera. A total of 67 bidirectional walking bouts were analyzed. No new benchmark datasets were introduced; the focus was on validating existing methods against a clinical standard.
Key Insights and Ideas
- 2D Pose Estimation Shines: 2D pose-estimation libraries (AlphaPose, Detectron, OpenPose) showed a stronger correlation with the Zeno walkway for most gait features compared to the 3D ROMP library. This suggests that for many gait parameters, the added complexity of 3D estimation might not be necessary or even beneficial with current monocular video approaches.
- Manual Heel-Strike Detection is Key (for some features): While automated heel-strike detection (ST-DBSCAN) showed promise, it struggled significantly with short, shuffling steps, which are common in PD. Manual annotation proved more reliable for features like the number of steps and cadence, leading to stronger correlations with the Zeno.
- Short Steps are a Challenge: The ST-DBSCAN algorithm's difficulty with short steps highlights a limitation for monitoring individuals with more severe PD or those experiencing freezing of gait. This necessitates manual annotation or further refinement of automated methods for such cases.
- Trends Over Absolute Values: The study emphasizes that while absolute values of gait features from video might not be directly interchangeable with Zeno measurements (due to different units and potential inaccuracies), the trends captured by video are highly correlated. This makes video analysis suitable for longitudinal monitoring.
- Camera Agnosticism: The robustness of the video analysis to variations in camera position and orientation is a significant finding, suggesting that the method could be applied in diverse, less controlled environments.
Key Results
- Moderate to strong positive correlations were found between Zeno and video-based gait features for number of steps, cadence, and mean/CV of step width when using 2D pose-estimation libraries and manual heel-strike detection.
- Correlations for the number of steps and cadence were significantly stronger with manual heel-strike annotation compared to automated ST-DBSCAN.
- The 3D ROMP library generally showed weaker correlations compared to the 2D libraries, particularly for gait features requiring depth information (speed, step length).
- A 20-step threshold was identified as a practical cut-off for the ST-DBSCAN method to maintain reasonable accuracy, excluding walks with an average step length of ~10.8 cm.
Limitations and Future Work
- Short Step Detection: The ST-DBSCAN algorithm's inability to reliably detect short, shuffling steps is a major limitation, requiring manual annotation for individuals with more severe PD.
- Absolute Value Discrepancy: Gait features calculated from video are not directly interchangeable with Zeno measurements due to differences in units and calculation methods.
- 3D Gait Features: While ROMP was explored for 3D gait features (speed, step length), its performance was weaker, suggesting challenges in accurately tracking anterior-posterior joint movement from monocular video.
Future work should focus on:
- Developing more robust automated heel-strike detection algorithms that can handle short and shuffling steps.
- Improving the estimation of 3D gait parameters from standard video.
- Exploring end-to-end machine learning models that can predict gait features directly from video without relying on intermediate joint trajectory extraction, potentially improving accuracy and incorporating depth information.
Summary updated