A Wearable Motion Capture Dataset for Gait Analysis Using IMUs and Shank-Mounted Egocentric Cameras

Participants

Ten participants (six male, four female; age: 23.9 ± 2.91 years; weight: 63.41 ± 6.81 kg; height: 1.65 ± 0.06 m) were voluntarily recruited from the local university community for this study. All participants were free from neurological or other medical conditions that could affect data collection.

The study was approved by the University of Central Florida’s Institutional Review Board (approved IRB number STUDY00002011), and informed written consent was obtained from each participant prior to the experiments. Participants consented both to participation in the study and to the sharing of their de-identified data for research purposes. To protect participant privacy, all data were de-identified prior to storage and distribution. Shank-mounted cameras providing a lower-limb egocentric view were used and therefore did not capture identifiable facial information of the participants. Any individuals who incidentally appeared within the video frame were anonymized using face-blurring techniques prior to data release.

Demographic information, including age, weight, height, and leg length distributions, is presented as boxplots in Fig. 1.

Performed Tasks

Each participant performed a total of 14 trials across four locomotion modes: treadmill, overground, slope, and stairs. For each locomotion mode, participants completed multiple trials that varied by speed, walking type, or included repetitions. In this section, we provide detailed descriptions of the trial types associated with each locomotion mode (Fig. 2).

• Treadmill. The participants walked on an instrumented dual-belt treadmill (Tandam, AMTI, Watertown, MA) for approximately two minutes at four different speed conditions. The order of the speed conditions was randomized for each participant, and a rest period was provided between trials to minimize muscle fatigue. Non-dimensional walking velocities (slow, normal, fast, and very fast) were defined based on each participant’s leg length, measured from the greater trochanter to the ground²⁹. The actual treadmill speeds were calculated by multiplying the non-dimensional velocities by $\sqrt{g{L}_{\mathrm{leg}}}$, where g is the gravitational acceleration and L_leg is the participant’s leg length. The resulting treadmill speeds (mean ± SD across participants) were 0.73 ± 0.02 m/s (slow), 1.11 ± 0.03 m/s (normal), 1.49 ± 0.04 m/s (fast), and 1.87 ± 0.05 m/s (very fast).

• Overground. Participants walked over a 5-meter overground path at four self-selected speeds: slow, normal, fast, and very fast. Each trial began from a predefined starting position, with participants walking straight along the path to a designated endpoint, where they stopped and turned back. No specific instructions were given regarding turning direction (left or right) to encourage natural walking patterns and variability in rotations. Each speed condition lasted approximately two minutes. The resulting walking speeds across participants were 0.67 ± 0.07 m/s (slow), 0.79 ± 0.09 m/s (normal), 0.99 ± 0.07 m/s (fast), and 1.16 ± 0.12 m/s (very fast). In addition to these trials, participants also completed a round walking trial along a circular path for about three minutes with a mean speed of 0.87 ± 0.06 m/s. A trial involving dodging two obstacles (chairs with dimensions of 80 cm × 55 cm × 50 cm) was also performed, with a speed of 0.75 ± 0.06 m/s across participants in the overground condition. From the starting position, the participant first walked straight toward the first obstacle, then turned left to go around it. After passing the first obstacle, the participant proceeded straight toward the second obstacle, turned right to go around it, and returned to the starting point. This cycle was repeated continuously for approximately three minutes of data collection.

• Slope. Participants also performed both uphill and downhill walking on a 20% slope^30,31 at self-selected speeds of 0.70 ± 0.06 m/s (ascent) and 0.67 ± 0.05 m/s (descent). Each condition was repeated twice to enrich the dataset with additional slope-walking data, providing greater variability and supporting the development of data-driven approaches. A rest period was provided between trials to minimize fatigue and ensure consistent performance across repetitions.

• Stair. Each participant completed two trials of stair ascent and descent on a five-step staircase (27-inch width, 9-inch depth, and 7-inch rise per step) at self-selected speeds of 0.27 ± 0.06 m/s (ascent) and 0.29 ± 0.06 m/s (descent). Both the stair ascent and descent conditions were repeated twice to capture natural variability and provide additional data.

Subject-specific walking speeds for all locomotion conditions are summarized in Supplementary Tables S1 and S2.

Motion Capture

Thirty-four markers were placed on each participant’s body based on a modified Helen-Hayes marker set³². A twelve-camera optical motion capture system (Vero, Vicon, Oxford, UK), operating at a sampling rate of 100 Hz, was used to record 3D marker trajectories. The detailed positions and names of the markers are provided in Fig. 3. After data collection, Vicon Nexus software was used to process each marker’s trajectory data. Occasionally occurring marker trajectory gaps, swapped markers, and unlabeled marker trajectories were carefully reviewed and addressed. Short gaps were automatically filled with built-in algorithms in Vicon Nexus. For the longer marker gaps, we prioritize using a rigid body fill technique, which regards a body segment as a rigid body and determines the missing marker trajectories with other existing marker trajectories within the body segment. For markers that are unable to use a rigid body filling technique, we use a kinematic filling technique based on the movement and geometry of a system. For the markers who could not process using the aforementioned techniques, we used spline interpolation and carefully reviewed the filled gaps. In addition to the primary marker set, occasional supplemental markers (such as T_head, R_hip, and L_hip) were applied for specific participants to aid in scaling the musculoskeletal model. Importantly, these markers were not used in the inverse kinematics process for calculating joint kinematics during dynamic walking trials, ensuring consistency across participants in the kinematic analysis. Marker trajectories were initially recorded in the Vicon motion capture coordinate frame (X: medio-lateral, Y: anterior–posterior, Z: vertical). These trajectories were transformed to the OpenSim global coordinate system (X: anterior–posterior, Y: vertical, Z: medio-lateral) prior to inverse kinematics analysis. All joint kinematics are reported in the OpenSim coordinate system.

Video Data

Video data were collected at 30 frames per second using shank-mounted GoPro cameras (Hero7 Black, GoPro Inc., San Mateo, California). The cameras were started via a GoPro remote before starting the motion capture system. Because the camera and motion capture systems were not hardware-synchronized, manual synchronization was performed by identifying multiple heel strikes visible in the video and aligning them with corresponding heel strikes from the motion capture marker data (Fig. 4). After initial alignment using a clearly identifiable heel strike, synchronization was visually verified at additional heel strike events occurring mid-trial to confirm consistent temporal alignment and the absence of observable time drift. Since the cameras were mounted on the shank, heel strikes of either the left or right foot were visible and could be accurately matched to the marker-based heel strike events. During data collection, researchers occasionally entered the field of view during experimental procedures. The shank-mounted video was dynamic, and faces were generally unclear. Additionally, all researchers wore face coverings throughout the study. To further ensure privacy protection, automated face detection was performed using InsightFace, and Gaussian blurring was applied to detected face regions using OpenCV in Python to obscure any potentially visible facial features in the footage.

Histogram of Optical Flow Features

Histogram of optical flow (HOF) features were extracted from the original, unmodified video data—prior to any face blurring or bitrate adjustments—to ensure the preservation of motion fidelity. Dense optical flow was computed using the Farneback method implemented in OpenCV³³. The flow fields between consecutive frames were converted to polar coordinates (magnitude and angle), and magnitudes were binned into 18 direction bins to produce normalized histograms representing motion density. HOF features are not available for the first frame because there is no previous frame to compute motion. We duplicated the histogram from the second frame for the first frame to keep the data synchronized. The histograms were normalized by the total number of pixels (assuming full HD resolution of 1920 × 1080). The full code and configuration used for feature extraction are provided in the GitHub repository for transparency and reproducibility. We also provide the face-blurring script separately; note that the released videos have been preprocessed solely for privacy protection and not for feature extraction purposes.

IMU

Eight IMUs (Avanti wireless EMG/IMU, Delsys, Boston, MA) were worn by the participants during the experiment. IMU data were collected at a native sampling rate of approximately 148 Hz. To synchronize with the motion capture data, the raw IMU signals (three-axis linear acceleration and angular velocity) were downsampled to 100 Hz. A trigger signal was used to start the Delsys system simultaneously with the Vicon motion capture system. Data collection was manually stopped on both systems, and the endpoint was aligned by matching available motion capture data to ensure full synchronization between the IMU and motion capture recordings. IMU data are expressed in local sensor-fixed coordinate frames defined by the orientation of each device.

Musculoskeletal Modeling and Joint Kinematics

OpenSim³⁴, an open-source musculoskeletal modeling tool, was used to perform scaling and inverse kinematics (IK). A musculoskeletal model (gait2354Simbody) was scaled for each participant using a static pose trial with a custom marker set based on participant-specific anthropometric data. During scaling, we ensured that the root mean square (RMS) error of the markers remained below 3 cm for all participants. Following scaling, the IK tool was applied to generate joint kinematics data by minimizing the difference between the model’s virtual markers and the experimentally collected markers. During IK, the RMS error of markers was kept below 3 cm to ensure accurate kinematic tracking. Joint angles are reported in degrees following standard Tait-Bryan conventions in OpenSim. Detailed definitions of joint angle conventions are provided in Supplementary Table S3.

Experimental Protocol

Before starting data collection, the motion capture system was calibrated using the standard protocol to ensure accurate 3D marker tracking. All markers, sensors, and camera positions were thoroughly verified to ensure consistency across participants. Marker placement was validated through the Vicon Nexus interface to confirm full visibility within the camera system. During each trial, coordinators closely monitored marker visibility and participant performance to ensure that movements were completed as instructed. If significant marker loss or placement issues were observed, the trial was halted and repeated to maintain data quality. Additionally, after initiating the trial in the motion capture system, the coordinator ensured that the IMU system was started for each trial. IMU data was also visually plotted during the trial run to verify that data collection was proceeding correctly. If any triggering issues arose, the entire triggering system was checked again to confirm proper operation. After completing each trial, another operator verified that all markers and sensors remained in their original positions and were securely attached to ensure consistent data collection for subsequent trials.

Motion Capture data pre-processing Validation

After data collection, all marker trajectory data were carefully inspected in Vicon Nexus to identify and correct any swapped markers, labeling errors, or trajectory gaps. After completing all the processing, the full trial is manually reviewed again to make sure it has biomechanically plausible trajectories to ensure the quality of data.

Musculoskeletal Modeling and Joint Kinematics

Marker trajectories collected from the motion capture system were used to scale the musculoskeletal model and perform inverse kinematics (IK) to generate motion data. For each trial, we ensured that the root-mean-square (RMS) error for scaling was below 3 cm and the RMS error for IK was below 3 cm to achieve high-quality musculoskeletal modeling outcomes. After generating motion data through IK, it was visualized again to verify that the motions were anatomically accurate and temporally consistent. Upon successful verification, the motion data were considered ready for further downstream analysis.

Data Visualizations

Figure 6 presents lower-limb joint kinematics (hip flexion, knee, and ankle angles) across the gait cycle (%) for different locomotion modes: overground, treadmill, round walking, obstacle dodging, slope ascent and descent, and stair ascent and descent. Each subplot presents the mean joint angle trajectory with shaded regions indicating standard deviation, reflecting natural gait cycle and inter-participant variability. For overground and treadmill conditions, the plots are stratified by walking speed (slow, normal, fast, and very fast), demonstrating the expected modulation of joint kinematics as speed changes. The shapes of these patterns match well with what has been reported in the literature²³. While the exact ranges of joint angles may differ due to cohort-specific factors (e.g., participant characteristics, ramp angles, and stair heights), the overall trends align well with typical gait characteristics. Notable patterns include greater knee flexion during stair ascent and increased ankle dorsiflexion during slope descent, both of which are consistent with established gait biomechanics.

Locomotion Mode Classification

In this section, we demonstrate a practical use case of our wearable dataset for gait-cycle-level locomotion mode classification across eight conditions using IMU and wearable camera (HOF) features. Gait cycles are segmented via heel strikes, resulting in ~14,000 samples. Per gait cycle, statistical features are extracted from both IMU and HOF signals. For IMU, six statistical features (mean, SD, RMS, max, min, and range) are computed for each of the 48 channels (8 sensors × 6 channels), yielding 288 features. For HOF, three statistical features (mean, SD, and RMS) are computed for each of the 18 feature dimensions from both left and right cameras, yielding 108 features. In total, 396 features are extracted per gait cycle. As a non-wearable baseline, six features—bilateral hip, knee, and ankle range of motion (ROM)—are extracted from each gait cycle. A Random Forest classifier is used for locomotion mode classification and is evaluated using leave-one-subject-out cross-validation (LOSO-CV). Features are z-score normalized using training data only; full implementation details are available in the accompanying GitHub repository.

Figure 7 summarizes classification performance across the ROM baseline and features derived from wearable sensors (chance level: 12.5%). The ROM baseline achieves an accuracy of 53.3%, whereas IMU and HOF features yield 79.8% and 82.1%, respectively. The combined IMU+HOF feature set achieves the highest accuracy (88.7%), indicating complementary information between inertial and egocentric vision modalities. Per-class results show that treadmill walking achieves 100% accuracy, while straight overground walking achieves 98% accuracy. Slope and stair conditions achieve accuracies in the range of 74–88%, potentially reflecting increased variability in movement patterns. In contrast, round and obstacle walking result in lower accuracies (64–65%), with misclassifications primarily occurring with standard overground gait and between these two conditions (Fig. 7B).

Kinematic Validation

To evaluate biomechanical consistency, joint kinematics are examined across treadmill and overground walking conditions at multiple speeds. Statistical comparisons of joint ROM are performed separately for each condition using the Friedman test, with Kendall’s W reported as an effect size. When the Friedman test is significant, post-hoc pairwise comparisons are conducted using Wilcoxon signed-rank tests with Bonferroni correction.

Across treadmill and overground walking conditions, sagittal-plane joint ROM shows consistent and physiologically plausible patterns across walking speeds (Fig. 8). For treadmill walking, hip, knee, and ankle ROM increase progressively with speed for both limbs, with hip flexion showing the largest relative increase (e.g., ~43^° to 60^° on the right side). Knee ROM shows moderate increases (approximately 65^° to 71^°), while ankle ROM increases from ~27^° to 35^°. These trends are statistically significant across all joints (p < 0.001), with moderate to strong effect sizes (Kendall’s W = 0.51–1.00). For overground walking, hip and ankle ROM generally increase with speed (e.g., hip: ~40^° to 49^°; ankle: ~26^° to 31^°), whereas knee ROM exhibits comparatively limited variation (approximately 62^° to 66^°) and is not statistically significant (p > 0.12). Overall, these results are consistent with prior studies reporting that joint kinematics vary with walking speed in healthy adults³⁶. Additionally, mean ROM values are closely matched between the left and right limbs across all joints and walking conditions, consistent with the expected bilateral symmetry of healthy gait (Fig. 8).

A direct comparison between treadmill and overground conditions suggests modestly higher ROM values during treadmill walking, particularly at higher speeds. While prior meta-analyses report that kinematics are broadly comparable between treadmill and overground walking despite some kinetic differences³⁷, the clearer speed-dependent ROM trends observed during treadmill walking may reflect the more controlled treadmill protocol, where walking speeds were predefined across a wider range. In contrast, the narrower overground speed range (0.67-1.16 m/s) — reflecting self-selected pacing on a short walking path involving turning and start-stop transitions — may reduce the separation of joint ROM across speed conditions, resulting in less pronounced speed-dependent trends during overground walking.

Limitations

The dataset was collected from a homogeneous group of healthy young adults in a controlled laboratory setting, which may limit generalizability to clinical populations and real-world conditions. Cameras were mounted exclusively on the shank at a fixed mounting configuration, which may restrict the field of view and does not capture broader environmental context available from alternative mounting positions.

File Types and Usage

The dataset includes multiple file types: .csv, .osim, .trc, .mot, and .mp4. Their typical uses are described below:

• .csv files can be opened using standard spreadsheet software and can be computationally analyzed using widely used data analysis frameworks. These files contain IMU data, kinematics, and HoF features.

• .osim files are OpenSim musculoskeletal models (e.g., Participant_xx.osim) that can be opened using the OpenSim application’s “Open Model” option for visualization or simulation.

• .trc files contain marker trajectory data and can be opened in spreadsheet software or directly imported into OpenSim. When paired with the corresponding .osim model, these files can be used with OpenSim’s inverse kinematics (IK) tool to compute joint kinematics.

• .mot files store joint kinematics and other motion data. They can be viewed in spreadsheet software, but are primarily intended for biomechanical analysis in OpenSim. After loading a participant’s musculoskeletal model (Participant_xx.osim), motion files (e.g., Pxx_overground_slow.mot) from the Kinematics(mot) folder can be visualized using the “Load Motion” option in OpenSim.

• .mp4 files contain video recordings of the trials and can be played using standard media players or analyzed frame-by-frame using video editing software. Additionally, these files can be processed with our provided code (available on GitHub) to extract HOF features.

Additional guidance on using OpenSim tools is available via the official OpenSim tutorials (https://simtk-confluence.stanford.edu:8443/display/OpenSim/Tutorials).

Example Applications of the Dataset

The dataset can support a variety of research projects utilizing multimodal IMU and video data for the estimation and prediction of joint kinematics. IMUs placed on top of the shoes have been used to estimate lower extremity joint angles using the DeepBBWAENet model³⁸, with segmented gait cycles across different trials utilizing this dataset. In another work³⁹, the combination of IMU data and HOF features extracted from videos was used to estimate whole-body pose and to predict future poses using the novel deep learning model AttRNet. Additionally, the Dual-ED-Attention-FAM-Net model⁴⁰ was developed using this dataset along with another public outdoor dataset²⁸ to predict future lower extremity kinematics.

Beyond these published studies, this dataset enables additional potential applications such as

• Locomotion mode recognition from egocentric shank-mounted video or sensor fusion for the potential development of prosthetic control strategies.

• Development and benchmarking of deep learning models with multi-modal wearables for kinematics estimation

• Domain adaptation studies using multimodal data across different locomotion modes or environments.

• Gait event/phase detection from the wearable IMUs and cameras.

By releasing this dataset in its entirety, we aim to support further research in multimodal wearable systems for gait analysis.