# Modified Extended Kalman Filter and Long Short-Term Memory-Based Framework for Reliable Stride-Length Estimation Using Inertial Sensors

## TL;DR

This paper introduces a new framework combining sensor preprocessing and machine learning to improve the accuracy of estimating stride length during walking.

## Contribution

The novel contribution is integrating a modified Kalman filter with LSTM networks to enhance stride-length estimation from noisy inertial sensor data.

## Key findings

- The framework reduced stride-length estimation error from 29.78% to 7.77%.
- LSTM models achieved a Mean Absolute Error of 0.0376 and an R2 score of 0.7066.
- The method proved robust for wearable gait analysis in clinical and health monitoring contexts.

## Abstract

Gait analysis plays a critical role in assessing mobility and identifying risks such as frailty and falls, where accurate spatiotemporal measurements are essential for early intervention, particularly in aging populations and clinical screening contexts. However, robust gait characterization remains challenging due to noise contamination and variability in sensor-based signals. To address these limitations, this study presents a stride-length estimation framework formulated as a modified processing-and-estimation pipeline integrated with Long Short-Term Memory (LSTM) networks. The pipeline includes wavelet-based denoising and cubic-spline interpolation as front-end preprocessing, followed by a Kalman-filtering stage with dynamic gain regulation guided by acceleration zero-crossing events to mitigate transient errors around abrupt turning points. Experimental data were collected from twelve healthy participants (seven females, mean age: 26.76 ± 3.01 years; five males, mean age: 25.81 ± 1.63 years) walking at self-selected speeds on a treadmill, using both an inertial sensor-based gait monitoring system and a motion capture system as the ground-truth reference. The proposed framework demonstrated a substantial improvement in stride-length estimation accuracy, reducing the absolute mean error from 29.78% to 7.77% and the standard deviation from 20.31 to 7.17. Furthermore, the LSTM models trained on Modified EKF-preprocessed data achieved superior performance metrics, with a Mean Absolute Error (MAE) of 0.0376 and a coefficient of determination (R2) of 0.7066. These results highlight the effectiveness of combining Modified EKF preprocessing with LSTM learning to enhance stride-length estimation reliability. This integrated approach offers a robust, noise-resilient solution for wearable gait analysis, providing valuable insights for clinical diagnostics, rehabilitation monitoring, and health management applications.