Structuring Variability in Human Gait Datasets: A Covariate-Centered Taxonomy and Systematic Review of Image- and...

Area of Science:

• Computer Vision
• Machine Learning
• Biometrics

Background:

• Human gait datasets are crucial for developing and evaluating computer vision models.
• Current datasets are heterogeneous, lacking consistent reporting of acquisition conditions, user variability, and sensor configurations.
• This heterogeneity limits reproducibility and cross-dataset comparability.

Purpose of the Study:

• To propose a standardized, covariate-centered, modality-agnostic taxonomy for human gait datasets.
• To enable consistent characterization of datasets by structuring variability across scene, user, and sensor factors.
• To identify systematic biases in current gait dataset design.

Main Methods:

• A systematic review protocol aligned with PRISMA 2020 was followed.
• 47 publicly available image- and depth-based human gait datasets were analyzed.
• A covariate-centered taxonomy (A-R) was applied for consistent characterization.

Main Results:

• The proposed taxonomy successfully structured variability across scene, user, and sensor levels.
• Quantitative analysis revealed systematic biases in covariate coverage across the analyzed datasets.
• The framework bridges differences across healthcare, biometric, and attribute-recognition domains.

Conclusions:

• The developed taxonomy provides a standardized framework for characterizing human gait datasets.
• This standardization enhances reproducibility and principled cross-dataset comparability.
• The findings highlight critical biases in current gait dataset design, informing future data collection efforts.