Title: Comparison of machine learning algorithms for an HAR problem using rotation times series data
Authors: Raphael Brard - University of Nantes (France) [presenting]
Lise Bellanger - Nantes University (France)
Aymeric Stamm - CNRS - Laboratoire de Mathematiques Jean Leray (France)
Pierre Drouin - Laboratoire de mathematiques Jean Leray - Nantes / UmanIT (France)
Laurent Chevreuil - UmanIT (France)
Fanny DOISTAU - UmanIT (France)
Abstract: In the last decade, there has been a growing interest in human activity recognition (HAR) algorithms based on inertial sensor data. The company UmanIT and the Department of Mathematics Jean Leray in Nantes have developed a solution for facilitating gait analysis on straight walk phases. To facilitate its use in real-life situations, we implemented and compared different machine learning algorithms (Support Vector Machine, Decision Tree, k-nearest neighbors and logistic regression) associated with post-treatment aiming at reducing false positive detections from a human daily recording. The data was collected by a motion sensor in the form of a unit quaternion time-series recording the hip rotation over time. This time series was then transformed into a real-valued time series of geodesic distances between consecutive quaternions. Moving average and moving variance versions of this time series were fed to the machine learning algorithms in order to train, tune and test our models. To compare the different models, we used metrics to assess the classification performances (AUC, accuracy) as well as metrics to assess change point detection capability and computation time. SVM stood out in terms of performance while decision trees led to the best compromise between performances and computation time.