Human Activity Dataset Generation - Smartphone-Based Human Activity Recognition - page 66

The processed inertial signals and the labels file (Sect. 4.3.1 ) were used for
the extraction and labeling of activity windows. The windowing process has been
approached in two ways. In the first method (

W 1 ), we worked with the start and
end times of each activity segment label and the signals defined within this time
region. We partitioned them into fixed-width sliding windows and assigned them the
corresponding activity label.
Inmethod two ( W 2 ), we first divided the entire inertial signal into sliding windows

(e.g. from the sequence of all the experiment activities) and then we assigned a label
to each window using the linked activity segments. This suggests that sometimes
uncertainty situations may appear. In particular, whenmore than one activity segment
label from the ground truth overlaps a particular window. To solve this, we defined
some conditions to choose the activity label to represent each window. In a window
sample, the time length of the involved labels was measured and it was chosen the
one that lasted longer.

W 1 but it is more
realistic because it takes into account that activities happen sequentially. It is more
useful for its application in online recognition systems where transitions between
activities occur. In the following sections we will be specific about which method is
used in the generation of the activity datasets.

W 2 is slightly harder to implement than

4.3.3 Feature Mapping and Dataset Generation

A reduced representation composed of features with relevant activity information can
be obtained from the activity windows in the time domain. These windows are also
transformed into frequency domain with the Discrete Fourier Transform (DFT) using
a real-valued FFT algorithm (Duhamel and Vetterli 1990 ;Ho 2004 ). In this work, we
extract features from these two domains. They include standard measures that have
already been proposed for HAR in several works (Yang et al. 2008 ) such as the mean,
correlation between signal pairs, Signal Magnitude Area (SMA) and autoregression
coefficients (Khan et al. 2010 ), energy of different frequency bands (Samà 2013 ). We
also include original measures such as frequency spectrum skewness and kurtosis,
and angles between triaxial signals. These measures are applied to the processed
accelerometer signals as well as the ones from the gyroscope (Table 4.3 ). Therefore,
considering the amount of signals involved, the generated number of features can
largely increase. Table 4.4 shows the selected signals and indicates the domains from
which features were extracted.
Table 4.5 shows the measures applied to the signals for generating the datasets
alongwith their formulation over the window signal s of length N . Numerical indexes
set under the s indicate one of the three possible axis x , y or z . A total of 561 features
were extracted to describe each activity window. Some of these features are well-
know and their estimation is straightforward. These are the ones in the top box of
the table. Others, on the other hand, are here introduced for clarification: