Human Activity Recognition Using Radial Basis Function Neural Network Trained via a Minimization of Localized Generalization Error

Abstract

Human activity recognition is a crucial component of applications in the areas of pervasive computing in healthcare. Human activity recognition approaches which adopt a data-driven approach are challenged by handling uncertainty in the data. These uncertainties arise due to sensor unreliability, natural noise and variance introduced by those performing the underlying activities. In this paper we propose an approach to human activity recognition based on Radial Basis Function Neural Networks (RBFNN) trained via a minimization of Localized Generalization Error in an effort to minimize the effects of uncertainty in the data. The proposed approach minimizes the generalization error taking into consideration both the training error and the stochastic sensitivity measure, which subsequently results in an improved generalization capability and improved tolerance to the uncertainty in the data. The approached developed was evaluated using data collected from the IESim smart environment simulation tool. Eleven activities were performed in a simulated environment, with uncertainty in the data stemming from user variance in completing the activities and the sensor placements in the environment. Classification accuracy of 98.86% was achieved demonstrating that the proposed RBFNN approach is robust to minor differences in unseen samples, many of which are caused by data uncertainty, following training which offers good generalization capability.