| MS Seminar

Name of the Speaker: Mr. Shrihari Sabu Nair (EE22S061)
Guide: Dr. Mohanasankar S
Online meeting link: http://meet.google.com/djy-sybm-iov
Date/Time: 15th May 2025 (Thursday), 2:30 PM
Title: A Computationally Efficient Framework for Vertical Oscillation Estimation via a Synergistic Signal Processing and Machine Learning Paradigm.

Abstract :

Monitoring Vertical Oscillation (VO) is essential for understanding gait mechanics, enhancing locomotor efficiency, and preventing injury. While wearable devices equipped with Inertial Measurement Units (IMUs) enable real-time VO tracking, their reliance on multiple sensors significantly increases power consumption, thereby limiting device longevity. This study proposes a computationally efficient framework for VO estimation using a single 3-axis accelerometer embedded in the chest-mounted sensor. The method employs a hybrid pipeline that combines classical signal processing techniques with machine learning models to enable accurate and real-time VO estimation under both constrained (treadmill) and unconstrained (field) conditions. Datawerecollected from30participants running on a treadmill, and vertical displacement was computed from the acceleration signal through a series of preprocessing steps, including RMS normalization, offset compensation, and detrended double integration. From the derived signals, relevant features such as ascent/descent amplitudes, durations, and velocity-area metrics were extracted to serve as inputs to multiple regression models. These models included Multiple Linear Regression (MLR), Random Forest, Support Vector Regression (SVR), Gradient Boosting Regressor (GBR), and Multi-Layer Perceptron (MLP).

Among these models, MLR achieved a mean absolute error (MAE) of 0.65 cm, mean squared error (MSE) of 0.75 cm2, and a Pearson correlation coefficient of 0.87 under treadmill conditions. When validated on an external dataset of six participants running in an open-field environment, the model retained strong performance (MAE: 0.78,cm), demonstrating its generalizability despite being trained solely on treadmill data. Other models such as SVR, GBR, and MLP demonstrated marginally better accuracy (MAE < 0.63 cm), especially in higher VO ranges; however, they were not selected due to their increased computational complexity and potential overhead during deployment. Giventhecloseperformancemarginsandconsideringdeploymentfeasibilityonembedded systems, MLR was selected as the preferred model due to its simplicity, interpretability, and computational efficiency. The proposed hybrid framework thus provides a power efficient and scalable solution for real-time VO monitoring, with applications in sports science, rehabilitation, and wearable health systems.