Year: 2017
Type: dissertation
Abstract: Human Activity Recognition (HAR) is a research topic with a relevant interest \nin the machine learning community. Understanding the activities that a person \nis performing and the context where they perform them has a huge importance \nin multiple applications, including medical research, security or patient monitoring. \nThe improvement of the smart-phones and inertial sensors technologies has \nlead to the implementation of activity recognition systems based on these devices, \neither by themselves or combining their information with other sensors. Since \nhumans perform their daily activities sequentially in a specific order, there exist \nsome temporal information in the physical activities that characterize the different \nhuman behaviour patterns. However, the most popular approach in HAR is to assume \nthat the data is conditionally independent, segmenting the data in different \nwindows and extracting the most relevant features from each segment. \n \nIn this thesis we employ the temporal information explicitly, where the raw data \nprovided by the wearable sensors is fed to the training models. Thus, we study \nhow to perform a Markov modelling implementation of a long-term monitoring \nHAR system with wearable sensors, and we address the existing open problems \narising while processing and training the data, combining different sensors and \nperforming the long-term monitoring with battery powered devices. \nEmploying directly the signals from the sensors to perform the recognition can \nlead to problems due to misplacements of the sensors on the body. We propose an \norientation correction algorithm based on quaternions to process the signals and \nfind a common frame reference for all of them independently on the position of the \nsensors or their orientation. This algorithm allows for a better activity recognition \nwhen feed to the classification algorithm when compared with similar approaches, \nand the quaternion transformations allow for a faster implementation. \nOne of the most popular algorithms to model time series data are Hidden \nMarkov Models (HMMs) and the training of the parameters of the model is performed \nusing the Baum-Welch algorithm. However, this algorithm converges to \n \nlocal maxima and the multiple initializations needed to avoid them makes it computationally expensive for large datasets. We propose employing the theory of \nspectral learning to develop a discriminative HMM that avoids the problems of \nthe Baum-Welch algorithm, outperforming it in both complexity and computational \ncost. \nWhen we implement a HAR system with several sensors, we need to consider \nhow to perform the combination of the information provided by them. Data fusion \ncan be performed either at signal level or at classification level. When performed \nat classification level, the usual approach is to combine the decisions of multiple \nclassifiers on the body to obtain the performed activities. However, in the simple \ncase with two classifiers, which can be a practical implementation of a HAR \nsystem, the combination reduces to selecting the most discriminative sensor, and \nno performance improvement is obtained against the single sensor implementation. \nIn this thesis, we propose to employ the soft-outputs of the classifiers in \nthe combination and we develop a method that considers the Markovian structure \nof the ground truth to capture the dynamics of the activities. We will show \nthat this method improves the recognition of the activities with respect to other \ncombination methods and with respect to the signal fusion case. \n \n \nFinally, in long-term monitoring HAR systems with wearable sensors we need \nto address the energy efficiency problem that is inherent to battery powered devices. \nThe most common approach to improve the energy efficiency of such devices \nis to reduce the amount of data acquired by the wearable sensors. In that sense, \nwe introduce a general framework for the energy efficiency of a system with multiple \nsensors under several energy restrictions. We propose a sensing strategy to \noptimize the temporal data acquisition based on computing the uncertainty of \nthe activities given the data and adapt the acquisition actively. Furthermore, we \ndevelop a sensor selection algorithm based on Bayesian Experimental Design to \nobtain the best configuration of sensors that performs the activity recognition accurately, allowing for a further improvement on the energy efficiency by limiting \nthe number of sensors employed in the acquisition.
Author: Nazabal Renteria, Alfredo
Institution: Universidad Carlos III de Madrid
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Keywords: human activity recognition, Markov model, hidden Markov models, spectral learning, Baum-Welch algorithm, sensor fusion, inertial sensors, wearable sensors, long-term monitoring, battery-powered devices, patient monitoring, smartphones, local maxima avoidance, computational complexity reduction
Open Access status: green
OpenAlex ID: W7011316477
