*2.3. Data Analysis*

#### 2.3.1. Pre-Processing

Raw marker trajectories were interpolated using Woltring gap filling [21] by Nexus (Vicon Nexus, Oxford, UK). Raw markerless video data were pre-processed by Theia3D (Theia3Dv2022.1.0.2309, Theia Markerless, Inc., Kingston, ON, Canada), where the default IK solution was used to estimate the 3D pose [18]. The lower body kinematic chain has six degrees-of-freedom (DOF) at the pelvis, three DOF at the hip, three DOF at the knee, and six DOF at the ankle. The kinematic and ground reaction force data were filtered through Nexus using a low-pass, zero-lag, 4-order Butterworth filter with cut-off frequencies of 10 Hz and 50 Hz [22], respectively.

#### 2.3.2. Visual3D Analyses

The pre-processed right lower extremity data were further analyzed using Visual3D (Preview v2022.06.02, C-Motion, Inc., Germantown, MD, USA).

The same Visual3D 6DOF algorithms and IK constraints for segments were adapted for both systems. IK constraints were set as six DOF at the pelvis, three at the hip, three at the knee, and six at the ankle. For the MB data, the human body was modeled by four linked segments (foot, leg, thigh, pelvis), in which a second kinematic-only foot was created as a virtual foot for kinematic estimations [23]. The segment mass estimations were based on Dempster's regression equation [24], and inertia properties were computed based on segments as geometrical shapes [25]. The hip joint center was estimated using the method proposed by Bell et al. [26]. Still, the knee and ankle joint centers were estimated using midpoints between external landmarks of the corresponding segment. The anatomical coordinate systems of segments were determined from the static calibration trial. The vertical axis was defined in the direction from distal to proximal joint center, while the anterior–posterior axis was defined as being perpendicular to the vertical axis with no mediolateral component. The third axis was the cross product of the vertical and anterior– posterior axis [27]. The model was automatically created for the ML data based on the deep learning algorithm and segment properties such as segment mass, location of the center of mass, and joint center positions were generated accordingly [19].

Resolved into the proximal coordinate system for both MB and ML data, joint angles and kinetic parameters in the sagittal plane were further calculated. The proximal segment was used as the reference when calculating joint moment and power. Internal joint moments and powers were obtained by applying Newton-Euler methods [1,28], where hip and knee extensor and ankle plantar–flexor moments were assigned to be positive. Positive power values indicated energy production through concentric muscular contractions [1].

Force-based gait events were used to identify stride cycles, in which the force threshold was set at 50 N [29]. The stride cycle was defined as two consecutive right heel contacts. The duration of each stride cycle was scaled to 101 data points.

#### 2.3.3. Discrete Measurements

The dependent variables were extracted from the last 10 strides from both MB and ML systems. Within each stride, various positive and negative peak values (depending on joint action) in the sagittal plane were identified on moment and power profiles of the hip, knee, and ankle joints, and the relative times to the peak values were included. Presented in Figure 1, peak moments of the hip (top panel) were extension moment in the early stance phase (HM1), flexion moment in the stance–swing transition phase (HM2), and extension moment at the end of the swing phase (HM3); for the knee (middle panel), the extension moment in the early stance phase (KM1), and flexion moment at the end of the swing phase (KM2); for the ankle (bottom panel), extension moment in the stance phase (AM1). Presented in Figure 2, peak powers of the hip (top panel) were absorption power in the middle of the stance phase (HP1), the production powers in the early swing phase (HP2), and at the end of the swing phase (HP3); for the knee (middle panel), absorption powers in the early stance phase (KP1), in the early swing phase (KP3), and at the end of the swing phase (KP4), and production power in the middle of the stance phase (KP2); for the ankle (bottom panel), absorption power in the early stance phase (AP1), and the production power at the end of the stance phase (AP2). –

**Figure 1.** Labeled here are the outcome variables used to quantify the differences in the lower extremity joint moments of the hip (**top panel**), knee (**middle panel**), and ankle (**bottom panel**) (denoted under the ensemble moment curve estimated by marker-based (MB) (red) and markerless (ML) (green) motion capture systems). Joint moments were scaled to participants' body mass.

**Figure 2.** Labeled here are the outcome variables used to quantify the differences in the lower extremity joint powers of the hip (**top panel**), knee (**middle panel**), and ankle (**bottom panel**) (denoted under the ensemble power curve estimated by marker-based (MB) (red) and markerless (ML) (green) motion capture systems). Joint powers were scaled to participants' body mass.
