Understanding how underlying systems of the human body interact to produce motion in normal or pathological states requires appropriate measurement, description, and analysis (Lu and Chang, 2012, Winter, 2009). Often direct measurements are difficult, or not possible, and musculoskeletal modeling could help understand the interaction between the different elements and the resulting movement biomechanics both in healthy and pathological conditions (Donno et al., 2022). However, the development of such models requires some in vivo measurements. An ideal measurement system should be non-invasive, minimally restrictive, allow for measurements in a natural environment, and have a sufficiently large field of view (Mündermann et al., 2006). Optical marker-based systems typically have millimeter level accuracy (Topley and Richards, 2020) but several limitations reduce their validity and ease of implementation in practice. Alternative systems using inertial measurement units (IMUs) (Zago et al., 2018) or integrated RGB images and depth sensors (RGB-D) (Eltoukhy et al., 2017, Oh et al., 2020) have been proposed, both approaches however have their own set of limitations including depth of view, integration drift, sampling rate, and accuracy concerns (Guffanti et al., 2020, Kumarasiri et al., 2018, Razavian et al., 2019, Yeung et al., 2021). Therefore, there is a need for innovative systems which encompass methods that overcome the above limitations.

Markerless motion capture methods using concepts in deep learning, particularly convolutional neural networks (CNNs) (Goodfellow et al., 2016), have outperformed classic computer vision algorithms on various tasks (Krizhevsky et al., 2012, Ren et al., 2015, Shelhamer et al., 2017). Importantly, CNNs have become an essential component for body part estimation in two-dimensional (2D) and three-dimensional (3D) human pose estimation (HPE) algorithms from RGB inputs (He et al., 2016, Huang et al., 2017, Newell et al.., 2016; Sandler et al., 2018, Tan and Le, 2019). Various monocular and multi-view approaches exist but the latter tend to perform better when evaluated on benchmark datasets (Desmarais et al., 2021).

Few studies have been conducted on the validity of available multi-view markerless systems outside of benchmark datasets, and even fewer have targeted 3D kinematics during gait. Several studies have compared gait spatiotemporal variables between marker-based and markerless approaches (Kanko et al., 2021, Moro et al., 2022) or markerless compared to GaitRite (Kanko et al., 2021, Lonini et al., 2022) and sensor-based (Mehdizadeh et al., 2021) methods. Previous results for the KinaTrax (KinaTrax Inc., Boca Raton, FL) markerless system showed excellent agreement between most spatial parameters calculated using coordinate-based gait events (Zeni et al., 2008) compared to those calculated from marker-based tracking and force plate events (Ripic et al., 2022). Worse agreement was found in temporal parameters. This prior error between approaches was influenced by inconsistencies between models and potentially due to coordinate-based gait events (Ripic et al., 2022) without the explicit comparison of other gait event methods. However, these two limitations are not independent from each other, and further work was warranted to explore the effect of gait event methods and their input variables. Importantly, only the ankle joint center was available in the previous markerless model to determine heel-strike whereas the heel marker was used in the marker-based model. Since the gait events are determined from peak positions of the proximal and distal foot segments relative to the pelvis in the anteroposterior direction, differences in input values from the two models could largely influence the event timing and subsequent parameters derived from these events. Additionally, the original work was done in only younger individuals, therefore performance of the markerless system in older adults and other groups remained unexplored.

The purpose of this study was to evaluate spatiotemporal parameters calculated with the latest version of the KinaTrax markerless model, which was retrained to include keypoints on the calcanei to improve consistency with the marker-based model. The project also aimed to address prior issues in gait event and sampling methods. We hypothesized that spatiotemporal parameters would show excellent agreement and consistency, nearly perfect concordance, and small bias and limits of agreement compared to a marker-based system. Additionally, better agreement was expected when comparing parameters calculated from velocity-based gait events.