Human Activity Dataset Generation - Smartphone-Based Human Activity Recognition - page 69

Table 4.6 Dataset partitions generated from the HAR experiment
Name

FP a Acc Gyro Time Freq d BAs PTs Sampling

- - 17 6 0 D 1 W 1

D 2

561 6 0

W 1

-

D 2 T

- 272 6 0

W 1

-

561 6 6 - D 3 W 2

D 3 T

- 272 6 6

W 2

-

a FP Fixed-Point

planned to provide a balanced number of samples per each BA, however during the
extraction of windows samples, classes resulted with a slightly different number of
activities between them producing a nearly-balanced set. We preserved all the sam-
ples and avoided using balancing methods such as homogeneous multi-partitioning
approaches (Aupetit 2009 ). Instead we worked with the parameters of the ML algo-
rithms used to balance the data when required.
In this work, different partitions of the dataset are used for specific purposes.
Therefore, in order to simplify the understanding in future sections, we collect in
Table 4.6 these partitions as well as their characteristics including type of sensors,
feature domains, activities, number of features and window sampling method. They
have been named with the symbol followed by a number which represents a

chronological order. We started with a simple version of the dataset for research
(

D

D 1 ). It has a limited number of features (17) which are extracted from accelerometer
data and it has fixed-point number representation. Following this, we added the
gyroscope and extended the number of features in the time and frequency domain
( D 2 ). Moreover, as we will see in Chap. 6 , we also consider the use of only time

domain features and we make distinction of this particular subset as

D 2 T . Finally the

D 3 and D 3 T partitions include PT information and modifies the window sampling

method required for the implementation of an online HAR system in Chap. 7 .

4.4 Results

In this section the generated HAR dataset is tested in order to verify its usability. This
is done in two ways. First, the inertial data are validated using various state-of-the-art
classification algorithms. Learned models from training data are used to predict new
samples and evaluate the algorithms' performance with the HAR test data. Second,
we present the results of a HAR competition that was organized in order to encourage
external researchers to find novel ML solutions to the same recognition problem.

4.4.1 Dataset Validation

A series of experiments were conducted for data validation. We employed some
of the most well-known ML algorithms as described in Sect. 2.5.2 . The Waikato