Background and Motivation

Human activity recognition $($HAR$)$ is the process of automatically recognising when a person is engaging in certain activities $($e$.$g$.$ running, walking, playing football$).$

A new smartphone app for HAR $($with a focus on walking$-$related activities$)$ is being developed.

The primary aim of this analysis is to distinguish between three different directions that a person is walking $($forwards$,$ backwards, circular motion$-$ see Figure $1).$

Feature engineering$/$extraction is the process of producing measurable characteristics for pattern recognition problems. Feature engineering is often utilised in HAR$-$computing and selecting meaningful summaries of the smartphone data $($as opposed to using the raw time series data$)$ for the goal of activity classification $($Straczkiewicz & Onnela, $2021).$

A secondary aim of this analysis is to compute and identify features that are important for distinguishing between walking$-$related activities.

The results of this analysis will be shared with the developers of the smartphone app.

Figure $1$: Visual representation of the walking$-$related activities.

The Data

Time Series Data

The dataset HAR$\_$walk.txt was collected in a laboratory setting, using the exact procedure outlined in Sikder & Nahid $(2021).$

To summarise, participants performed a walking$-$related activity for $3$ seconds, while recordings were made via a smartphone $($see Figure $2).$

Data pre$-$processing has also been performed using the exact procedure outlined in Sikder & Nahid $(2021).$

The dataset HAR$\_$walk.txt consists of accelerometer signals and gyroscope signals $($in each of the $x,y$ and $z$ directions$)$ recorded at $300$ time points. An example set of six time series is displayed in Figure $3.$

The dataset HAR$\_$walk.txt contains $900$ observations $(324$ walking forwards, $317$ walking backwards and $259$ walking in a circular motion$).$ The class labels are in column $1801($labelled "Type"$).$

Figure $2$ Diagram of the data collection process for the dataset $($resulting in six time series for each observation$).$

r$$gure s An examp$$e of the s$$x t$$me ser$$es Jrom acce$$erometer ana gyroscope s$$gna$$s of an observat$$on of a person wa$$ng forwards.

Feature Engineering

Feature engineering was performed to create new variables $($using the time series data in HAR$\_$walk.txt$)$ for the purpose of classification.

For each observation, the mean, standard deviation and minimum value for each of the six time series $($x$/$y$/$z$-$axis for Accelerometer$/$ Gyroscope signals$)$ was computed, resulting in $18$ engineered features.

A further two features $($labelled $"$V$19"$ and $"$V$20")$ were also engineered.

The dataset HAR$\_$walk$\_$FE$.$txt contains $20$ engineered features for the $900$ observations in the original dataset. The first column $("$Type$")$ contains the class labels $($i$.$e$.$ walking direction$).$ See Table $1$ for further details.

$\backslash$table$[[$Walking Direction,Class Label,Row Numbers$],[$Forward$,3,1$:$324$