A Reproducible Multi-Camera Framework for Synchronized Video Acquisition, Calibration, 3D Human Pose Reconstruction, and Markerless Gait Analysis

Background Quantitative gait analysis provides objective measures of human locomotion that are useful in biomechanics, rehabilitation, mobility assessment, and movement research. Temporal correspondence, camera calibration, cross-view keypoint association, triangulation, and downstream gait measurement can each introduce error, yet these stages are often evaluated separately or treated as hidden processing steps. Conventional three-dimensional motion-capture systems provide accurate measurements but require specialized laboratory infrastructure, calibrated capture volumes, reflective markers, and dedicated hardware. We developed a markerless system that combines synchronized video, camera calibration, multi-view pose estimation, triangulation, and gait analysis in a single acquisition and processing workflow. Objective We developed and evaluated a reproducible multi-camera system that treats synchronized video acquisition, geometric camera calibration, markerless two-dimensional pose estimation, metric three-dimensional reconstruction, and extraction of quantitative gait parameters. The current implementation uses two cameras and was designed so that additional viewpoints and inertial measurements can be incorporated without changing the basic acquisition workflow. Methods We recorded the same walking volume simultaneously with two USB cameras. We calibrated each camera to estimate intrinsic parameters and lens distortion using a ChArUco target, then estimated the relative camera geometry for multi-view reconstruction. Anatomical keypoints were detected independently in each view and paired across synchronized frames. Corresponding 2D observations were triangulated to obtain 3D joint coordinates. Confidence values, reprojection checks, outlier rejection, and temporal processing were retained as separate steps so that errors could be traced through the pipeline. Validation We evaluated the system at the camera-calibration, temporal-synchronization, geometric-reconstruction, human-pose-reconstruction, and gait-analysis levels. Calibration performance was quantified from reprojection residuals, temporal correspondence from recorded frame timestamps, and geometric reconstruction from the corridor-global calibration. A measured staircase provided an additional physical reference for vertical reconstruction. Level walking was used to demonstrate continuous 3D human-pose reconstruction and extraction of preliminary spatiotemporal gait variables. Conclusion The principal contribution is a reproducible and traceable measurement chain that connects synchronized multi-camera acquisition to markerless 3D locomotion analysis while preserving the intermediate information needed to identify calibration, synchronization, pose-estimation, and reconstruction errors. Rather than proposing a new calibration or pose-estimation algorithm, the framework integrates established components into a session-based system that can be reprocessed, audited, and extended to additional cameras and inertial sensors, also use in an more complex locomotor tasks.