Human Activity Dataset Generation - Smartphone-Based Human Activity Recognition - page 55

Chapter 4
Human Activity Dataset Generation
4.1 Introduction
Two of the vital elements required for developing our research in smartphone-based
HAR are: experimental data collection and dataset generation. In this chapter, we
describe a general overview of these elements organized in three main sections.
Section 4.2 describes all the aspects linked to data gathering, first by selecting the
appropriate smartphone for experimentation, and then by describing the protocol of
trials with volunteers. Moreover, Sect. 4.3 deals with the raw sensor data in order to
generate the dataset. For this task, we use signal conditioning techniques and select
a suitable set of features for data characterization. In Sect. 4.4 preliminary results
with the available data are presented. This includes data validation with different ML
algorithms that confirm their usability and data publication in a centralized repository.
We also describe an organized HAR competition with the obtained dataset in which
people were encouraged to propose their own solutions to the recognition problem.
Finally, the chapter is summarized in Sect. 4.5 .
4.2 Experimental Data Collection
In the HAR research framework, some benchmark datasets have been released to the
public domain providing experimental data with various inertial sensors. They pro-
vide a freely available source of data across different disciplines and researchers in
the field. For example, the Opportunity Project Roggen et al. ( 2010 ) has recorded a
set of ADL in a sensor-rich environment using 72 environmental and body sen-
sors. Similarly, other works have provided datasets such as Tapia et al. ( 2006 )
and Dernbach et al. ( 2012 ).
However, public smartphone inertial data for activity recognition is limited. For
this reason, we chose to make our own collection of data and also to make it available