A Random Forest Based Methodology for the Development of an Intelligent Classifier of Physical Activities

Abstract

An appropriate amount of Physical Activity (PA) has been demonstrated to be positive for health. However, most of the developed approaches to monitor PA do not consider people that require assistive devices for walking (ADW), in which gait patterns differ significantly from normal gait. In this work, a methodology is proposed to define a neural-network based classifier that uses the data from a sensorized crutch tip to detect the different PA. To achieve this, a series of features are obtained from the sensorized tip’s sensors, and a proper selection is carried out using a Random Forest approach. Based on the relative influence of each feature, a set of them are selected to define a Multi-Layer Perceptron neural network to classify four types of PA. Results show that using this procedure leads to appropriate classifiers.

This research was supported by the University of the Basque Country under grant number PIF18/067 and project number GIU19/045 (GV/EJ IT1381-19) and by the Ministerio de Ciencia e Innovación (MCI) under grant number DPI2017-82694-R (AEI/FEDER, UE).