Step Detection using SVM on NURVV Trackers

This paper introduces a Machine Learning approach (ML) for classifying step detection during human running activities. First, we use a signal processing strategy to label Inertial Measurement Unit (IMU) data (i.e. acceleration and angular speed) in terms of foot contact, ground vs air. This is done by performing Exploratory Data Analysis (EDA), that includes Principal Component Analysis (PCA) for interpretability, on a collection of IMU data sets obtained via multiple runners using a NURVV Run wearable device. Once we are in the presence of a supervised learning problem, by leveraging ML techniques - such as Support Vector Machine (SVM) - we can optimize models to detect if the foot is in the air or on the ground solely based on IMU data. Unlike in this first instance where we rely on signal processing, this algorithm is designed to not need any post-processing, i.e. if the embedded system has enough resources it should be able to run in real-time. Since the raw IMU data is affected by factors such as the position of trackers on the shoes, running speed, runner technique and terrain, a single model doesn't generalise well. Therefore, we implement an ensemble SVM model, that relies on the confidence that each separate SVM model has on the output of its own classification to extract, through hard-voting, the classification of the sample. We present promising initial results from applying this approach to unseen test data.