Markerless motion capture methodologies in biomechanics: From 2D images to 3D joint angles, a narrative review

Video-based markerless motion capture is rapidly emerging as a valuable tool in biomechanics research, particularly within sports science, ergonomics, and clinical evaluations. Markerless motion capture offers greater versatility for capturing movement outside of laboratory settings and with minimal setup. This technique involves a series of steps to derive 3D joint angles from 2D video images: 2D keypoint identification, triangulation, and inverse kinematics.

What methodologies are used at each step of the process from 2D images into 3D joint angles?

This narrative review critically analyses the methodologies employed at each step of the process based on 12 methodological articles of the literature. These articles are representative of typical commercial or research approaches currently available.

Most markerless approaches use OpenPose to identify the 2D keypoints, some of them correct their inconsistencies either in 2D or 3D or increase their number with augmented 3D landmarks. Few approaches use a confidence-weighted triangulation. Almost all of them perform multibody kinematics optimisation, classically tracking the 3D keypoints/landmarks, or maximising the confidence derived from all the images.

This process shares certain similarities with traditional marker-based motion capture. The methodologies are adapted to a limited number of keypoints with unclear anatomical definition and potential physically implausible trajectories. The methodologies are also adapted to incorporate the confidence information. Most adaptations rely on neural networks complementary to the human pose estimators. Markerless motion capture methodologies are close to what is commonly applied to marker-based data acquired with opto-electronic cameras but mostly rely on data-driven statistical inference. The evaluation is currently limited to comparison with marker-based kinematics, mainly on young adults without pathology.