Alfredo Nazábal Rentería, a PhD student in the Signal Processing Group of the University Carlos III de Madrid has defended his doctoral thesis titled «Markov Modelling on Human Activity Recognition» on September, 29th

• Title: “Markov Modelling on Human Activity Recognition»
• Advisor: Antonio Artés Rodríguez.
• Event Date: Friday, September 29, 2017, 11:00 am.
• Location: Sala Adroración de Miguel (Betancourt Building) Leganés Campus; Universidad Carlos III de Madrid.

Abstract

Human Activity Recognition (HAR) is a research topic with a relevant interest in the machine learning community. Understanding the activities that a person is performing and the context where they perform them has a huge importance in multiple applications, including medical research, security or military, among others. The improvement of the smartphones and inertial sensors technology has lead to the implementation of activity recognition systems based on these devices, either by themselves or combining their information with additional sensing devices. Since humans perform their daily activities sequentially in a specific order, information about the current physical activities depend on the previous ones, creating time sequences that characterize the different human behaviour patterns. However, the most popular approach in HAR is assuming that the data is conditionally independent, losing the temporal relation of the activities, and segmenting the data in different windows, extracting features from each segment that are relevant for the classification of the activities.

In this thesis we follow a different approach. We employ the raw signals provided by the wearable sensors directly, with no segmentation process, and we use them to feed classification algorithms. With typical sampling frequencies of wearable sensors ranging from ten to hundreds of Hertz, the dynamics of the activities are retained in the data, and current observations are closely related to the previous ones. Thus, we study how Markov modelling enhances the implementation of HAR systems with wearable sensors by maintaining this temporal relation. We address the existing open problems arising due to the definition of the HAR problem employed in this thesis, that is, the implementation of long-term monitoring systems which combine the raw signals from different wearable sensors.

First, the raw physical signals are extremely sensitive to the location of the sensors on the body and to the presence of misplacements. We propose an orientation correction algorithm that transforms the data provided by the sensors, returning it in the same reference system independently on the position of the sensors or their orientation. This algorithm allows for a better activity recognition by feeding the corrected data to the classification algorithms, when compared with similar approaches, and the quaternion transformations allow for a faster implementation.

Training the parameters of a Hidden Markov Model (HMM) is usually per- formed using the Baum-Welch algorithm. However, convergence is not guaranteed and multiple initializations are needed to avoid local maxima, and consequently, the algorithm becomes computationally expensive in large datasets. We propose employing spectral learning to train a discriminative HMM, avoiding these problems while simultaneously maintaining the performance of the classification.

Later, we address the problem of combining of a small number of classifiers during activity recognition, which is particularly relevant in HAR, since the number of sensors is usually reduced as much as possible. In the simplest case of two classifiers, which can be a practical implementation of a HAR system, the combination reduces to selecting the most discriminative sensor, and no performance improvement is obtained against a single sensor implementation.

In this thesis, we propose to employ the soft-outputs of the classifiers in the combination, and we develop a method that considers the Markovian structure of the ground truth to capture the dynamics of the activities. We show that this method improves the recognition of the activities with respect to other combination methods and with respect to combining the raw signals of the sensors directly.

Finally, we study the energy efficiency of wearable sensors in the implementation of a HAR system. The most common approach to improve the energy efficiency of such devices is to reduce the amount of data acquired by the wearable sensors. We propose a sensing strategy, that we call active sensing, which optimizes the data acquisition by computing the uncertainty of the activities given the data and actively adapting the acquisition. Furthermore, we develop a sensor selection algorithm based on Bayesian Experimental Design to obtain the best configuration of sensors that accurately performs the activity recognition, allowing for a further improvement of the energy efficiency by limiting the number of sensors employed in the acquisition.