E-Locus - Institutional Repository of the University of Crete

Home Search

Results - Details

Search command : Author="Μουχτάρης" And Author="Αθανάσιος"

Current Record: 3 of 29

[Add to Basket]

Identifier

000414781

Title

Feature selection architectures for human activity recognition

Alternative Title

Αρχιτεκτονικές επιλογής χαρακτηρισικών για την αναγνώριση ανθρώπινης δραστηριότητας

Author

Καραγιαννάκη, Αικατερίνη Ε.

Thesis advisor

Τσακαλίδης, Παναγιώτης

Reviewer

Μούχταρης, Αθανάσιος
Laerhoven, Kristof Van

Abstract

The continuous evolution of sensing technologies during the last decades has resulted in the production of massive streams of daily generated data for processing and interpretation. The performance of Human Activity Recognition (HAR) systems is directly affected by the volume of the produced raw signals due to the continuous need for rapid and accurate predictions, especially in the case of online/real-time installations. Feature Selection techniques, as an integral aspect of the Machine Learning/Data Mining pipeline, reduce the volume of the available raw streams and filter out redundancy, providing the classifiers with discriminative information about the available data. This thesis explores the effect of Feature Selection in the HAR domain and it involves three phases. During the first phase, we perform a benchmark study in order to assess the optimal parameters, experimental set-up, and suitable algorithms for incorporating Feature Selection in HAR applications. We evaluate the pipeline for Data Acquisition, Segmentation, Feature Extraction, and Feature Selection on two different HAR datasets. Our work demonstrates that the use of short windows during the segmentation stage results in better classification performance, since the involved data in each window is characterized by fewer class labels. Concerning Feature Selection, we highlight the effectiveness of unsupervised graph-based methods. The second phase of the thesis focuses on the transition into online HAR applications by incorporating the findings of the benchmark study into an online environment by means of an Android application. The Data Acquisition process is performed using smaller batches of data instead of large datasets, simulating a streaming service scenario. We explore different batch sizes and observe the performance of various Feature Selection algorithms for various data partitions, with respect to the available activities. We also perform a qualitative analysis on the selected features and we extract information about the main modalities that convey dominant features. We evaluate the online performance of the distinct application components in terms of execution time and we measure the overall energy requirements on the Android platform. Our experimental results highlight the contribution of short batches of data to the faster execution of the individual components, and the efficacy of graph-based techniques to select dominant attributes. The third and final phase of this thesis explores techniques and architectures for online feature-level fusion. We expand our Android Feature Selection library and include methods that merge data based on two approaches, an online Classification scheme and a Feature Selection architecture, both operating in two stages. The first stage of both approaches involves the execution of the Machine Learning pipeline on data originating from different sensor node locations. For the online Classification scheme, the selected features from each location are merged into a concatenated feature matrix for the further prediction of the underlying Human Activities on the Android device. The Feature Selection architecture proceeds into a second layer of feature selection by utilizing the concatenated matrix of features from different node locations, enriched with inter-location pairwise correlations. Our results highlight the efficacy of the Feature Selection architecture to provide high data compression with a minimal computation time overhead.

Language

English

Subject

Machine learning

Wearable devices

Μηχανική μάθηση

Φορητές συσκευές

Issue date

2018-03-23

Collection

School/Department--School of Sciences and Engineering--Department of Computer Science--Post-graduate theses

Type of Work--Post-graduate theses