Accurate stride length estimation is crucial for gait analysis which has diverse applications in positioning and clinical domains. In positioning systems, particularly in pedestrian dead reckoning (PDR) and inertial navigation, precise stride length estimation directly influences the accuracy of trajectory reconstruction and displacement measurement. It plays a vital role in enabling reliable indoor localization in global navigation satellite system (GNSS) denied environments, such as in hospitals, airports, and shopping malls. This study presents a novel method for stride length estimation using Ultra-Wideband (UWB) sensors mounted on the shins of the user. UWB distance measurements, captured at 10 Hz, were smoothed and segmented into individual gait cycles using inertial data to detect heel strike and toe-of events. Since UWB measurements reflect stride length at the shin level, motion capture data was used to calibrate these measurements to actual foot-level stride lengths via a linear regression model. The model was trained on data from three separate experimental sessions and evaluated using root mean square error (RMSE), coeficient of determination (R²), and paired t-tests. The calibrated UWB-based method demonstrated high agreement with motion capture reference data, achieving an RMSE of 0.057 m and R² of 0.897 on test data. Comparative analysis further showed a significant improvement over an IMU-only approach, with the UWB method yielding a lower mean error (0.036 m vs. 0.082 m). These findings support the viability of UWB sensors, when properly calibrated, as an efective and wearable solution for stride length estimation.
Stride length estimation is a crucial component of pedestrian positioning systems, enabling accurate
trajectory reconstruction and improved localization. While most existing methods [
Zero velocity updates (ZUPT) are widely used to reset velocity during stance phases, reducing drift
accumulation [
In this work, we propose a low-cost, drift-free alternative for stride length estimation using UltraWideband (UWB) sensors. UWB enables accurate time-of-flight (ToF) distance measurements through
CEUR
ceur-ws.org
short-duration radio pulses across a wide frequency spectrum. Its high temporal resolution, low power
consumption, and robustness in non-line-of-sight (NLOS) conditions make it highly suitable for wearable
applications in navigation, sports, and health monitoring. Unlike previous UWB-based methods that
rely on fixed infrastructure [
Human locomotion follows a cyclic sequence of movements, collectively referred to as the gait cycle
[
In this study, participants wore shoes equipped with a 9-axis Xsens DOT inertial measurement unit
(IMU) mounted on the heel of the left foot [
X-axis is oriented opposite to gravity, the Z-axis points outward from the foot, and the Y-axis completes
a right-handed system. The device recorded triaxial acceleration, angular velocity, and magnetic field
data at a sampling rate of 60 Hz. Data were transmitted via Bluetooth to a smartphone running the
Movella DOT app, which was used for data extraction. In parallel, Decawave DWM1001 Ultra-Wideband
(UWB) modules were mounted on each shin to capture inter-shank distances at a sampling frequency
of 10 Hz [
Stride length estimation in this work uses UWB ranging measurements along with inertial data, using OMC data as ground truth. The experimental setup is described in Section 3.1. Initially UWB data is preprocessed to reduce measurement noise and signal fluctuations (Section 4.1). All datasets are then temporally synchronized as explained in Section 4.2. Reference stride lengths are derived from highresolution 3D positional data obtained using retroreflective markers attached to each foot, following the methodology in Section 4.3. Then gait event detection and stride length estimation (Section ??) is performed by first identifying HS and TO events and then performing stride length estimation using both UWB and OMC data, as outlined in Section 4.4.
UWB sensors mounted on the shin record the Euclidean distance between the two sensor nodes at a
sampling rate of 10 Hz. To mitigate the impact of measurement noise and minor signal fluctuations,
the raw UWB signal is subjected to a preprocessing procedure described in Kumar et al. (2024) [
Due to independent acquisition systems, a temporal ofset exists between the UWB and OMC datasets. The OMC system provides 3D marker positions at 100 Hz, while the UWB system records inter-shin distances at 10 Hz. To enable synchronization, a comparable signal is derived from the OMC data by computing the Euclidean distance between markers attached to the shin-mounted UWB units (see Fig. 2), as given by: = √( − )2 + ( − )2 + ( − )2 (1) where ( , , ) and ( , ,
) represent the coordinates of the left and right UWB markers at time ,
respectively. Prior to cross-correlation, the OMC data is downsampled to 10 Hz to match the UWB data.
The temporal ofset is determined using cross-correlation analysis [
UWB signals alone cannot directly identify heel strike (HS) and toe-of (TO) events. Two approaches can address this limitation. The first uses prior knowledge of the user’s initial step (e.g., left or right foot), enabling inference of HS and TO from the UWB trajectory. The second approach, more reliable, leverages foot-mounted IMU data to detect HS events, which often align with peaks in the UWB signal. This correlation allows accurate identification of both HS and TO events—provided the IMU and UWB data are time-synchronized. Since Xsens DOT IMUs are timestamped using the smartphone’s clock, which is unsynchronized with the UWB system, synchronization is achieved via optical motion capture (OMC) data. A common event—the onset of walking—is used for alignment by analyzing vertical motion. The Z-axis heel marker position from OMC and gravity-aligned IMU acceleration are compared during an initial stationary phase. A threshold on the moving mean (window size WW) detects the first significant deviation, indicating the start of movement. The resulting temporal ofset aligns the IMU and OMC data. An example of this synchronization is shown in Fig. 4.
The OMC system is employed to generate ground truth stride length data. For stride length computation,
the heel marker placed directly on the Xsens DOT sensor mounted at the heel was primarily utilized.
Ground truth stride length is computed as the Euclidean distance between the 3D positions of the heel
marker at successive heel strikes. The stride length is calculated using the following formula:
=
√( 2 − 1)2 + ( 2 − 1)2 + ( 2 − 1)2
(2)
where ( 1, 1, 1) and ( 2, 2, 2) denote the 3D coordinates of the heel marker at the initial and
subsequent heel strike events, respectively. Heel strike events were manually annotated using files
obtained from the OMC system. These files were visualized and processed using Mokka, an open-source,
cross-platform software for biomechanical data analysis [
Following signal preprocessing, gait event detection is performed using inertial measurement units
(IMUs) attached to the feet. Heel strike (HS) and toe-of (TO) events are identified using the method
proposed by Kumar et al. (2024) [
Stride length is initially estimated at the shank level by summing the inter-sensor distances recorded at the TO of the current cycle and the HS of the subsequent cycle. As these UWB-derived measurements reflect movement at the shin rather than the foot, a calibration procedure is applied to align them with foot-level stride lengths obtained from a motion capture system. Specifically, a linear regression model is trained using paired data: UWB-based stride lengths as input and motion capture-based stride lengths as target output. The trained model is then used to calibrate the UWB estimates, and the performance of these calibrated stride lengths is evaluated against the ground truth.
Using the hardware setup explained in section 3.1, three separate experiments were conducted with a single participant. In each experiment, participant is instructed to progressively increase their step length relative to the preceding trial. A rectangular trajectory is marked in the OMC area with a rectangular box marking the starting and ending point of the experiment. In each experiment the participant remains stationary for 1 minute inside the marked rectangular box after which they start walking along the rectangular trajectory. During each experiment, participant completed two full laps of the rectangular path. Only data collected while participant was inside the motion capture area is retained for analysis; any data recorded outside this region is excluded. Furthermore, only straight-line walking segments were considered for stride length analysis. Steps taken during turns were omitted to minimize the influence of directional changes on stride length estimation accuracy. Details regarding the number of steps analyzed, and the mean stride length per experiment are summarized in Table 1.
To develop and evaluate the calibration model, stride length data from all three experimental sessions were partitioned into training (80%) and test (20%) sets. Stratified sampling was used to ensure that both sets reflected the full range of observed stride lengths, minimizing sampling bias and enhancing generalizability across diverse gait patterns. A linear regression model was trained using UWBderived stride lengths measured at the shin as input and the corresponding foot-level stride lengths from the optical motion capture (OMC) system as ground truth. To ensure robustness and prevent overfitting, 5-fold cross-validation was applied to the training data for model parameter estimation. Model performance was assessed on the independent test set using standard regression metrics. As shown in Fig. 6, the model achieved a root mean square error (RMSE) of 0.057 m and a coeficient of determination (R²) of 0.897, indicating that the model explained approximately 89.7% of the variance in the OMC-based reference stride lengths. A paired t-test comparing predicted and ground truth stride lengths yielded p = 0.115, indicating no statistically significant diference and suggesting that prediction errors were random rather than systematic. Detailed results across individual experimental sessions are presented in Table 2. Experiment 1 and 3 exhibited moderate agreement, with RMSE values of 0.053 m and 0.073 m, and R² values of 0.392 and 0.515, respectively. Corresponding t-tests (p = 0.068 and p = 0.244) confirmed the absence of significant bias, although the reduced R² values suggest variability in model performance. In contrast, Experiment 2 showed a negative R² (–1.727), indicating poor trend capture despite a relatively low RMSE of 0.047 m. The high p-value (0.591) suggests that predictions were not significantly diferent from ground truth but lacked consistent directional agreement. Overall, these results demonstrate the efectiveness of the proposed UWB-to-foot-level calibration model, while highlighting the need for further investigation into inter-experiment variability to improve robustness under varied gait conditions.
To further evaluate the accuracy of the calibrated UWB-based stride length estimation, a comparative
analysis was conducted against stride lengths derived from strap down inertial navigation approach
used by Suzuki et al. [
This study introduced a novel method for stride length estimation using shin-mounted Ultra-Wideband (UWB) sensors, integrated with inertial data for gait event detection and calibrated against motion capture data. A linear calibration model was developed to map UWB-derived stride lengths to footlevel reference values. Trained and validated on data from three experimental sessions, the model demonstrated strong performance with an RMSE of 0.057 m and R² of 0.897 on the test set. A paired t-test confirmed no significant diference between calibrated UWB estimates and motion capture ground truth, indicating high accuracy and consistency. Comparison with a conventional IMU-based method revealed a lower mean error for the UWB approach (0.036 m vs. 0.082 m), underscoring its potential as a robust alternative in GNSS-denied or constrained environments. Boxplot analyses further highlighted the reduced variability and improved reliability of the UWB-based estimates post-calibration. Future work will aim to generalize this method to diverse populations, including elderly and clinical cohorts with atypical gait. Additionally, the current linear calibration could be extended to nonlinear or machine learning-based models to better capture complex gait dynamics. Enhancements such as multi-sensor UWB configurations for stride symmetry analysis and real-time wearable implementation will also be explored to improve system scalability and practical deployment.
The authors thank Prof. Abhishek, Department of Aerospace Engineering, IIT Kanpur, for access to the motion capture lab in the Helicopter Building. They acknowledge the University of Melbourne and IIT Kanpur for the Melbourne India Postgraduate Academy Scholarship, as well as partial funding from the Government of India’s SPARC project (ID: P3146).
During the preparation of this work, the author(s) used X-GPT-4 and Gramby in order to: Grammar and spelling check. After using these tool(s)/service(s), the author(s) reviewed and edited the content as needed and take(s) full responsibility for the publication’s content.