14
A. Chatzimichail et al.
and meaning to the content of data coming from IoT devices and interlink related
data.
3.1 Modelling Sensors
Sensors embedded in devices that are attached on the human body or sensors directly
placed on the human body, are called wearable sensors [89]. The existence of such
sensors in mobile and wearable devices has led to their extensive use in activity
recognition and fall detection tasks. Accelerometers and gyroscopes are the most
popular ones, with accelerometers being the most effective in recognising activities
when used individually. Gyroscopes are also quite popular, however, they are mostly
used in combination with the accelerometers. Accelerometers are known to perform
well in recognizing activities in general but they are more successful in activities
with repetitive movement [89], since they measure a moving object’s magnitude and
direction. They usually fail to recognize similar activities when used individually
[90], thus it is more effective to use them along with other inertial sensors to improve
the performance of a human activity recognition system. Gyroscopes perform well
in recognizing an object’s orientation because they measure the rotation speed [91].
Gyroscopes are widely used in activity recognition studies, but most of the times
not as the only sensor. Fusion of such sensors, whether performed before or after the
classification algorithm, is found to improve the recognition rates of a human activity
recognition system, since one sensor may capture movements not well detected by the
other [92]. Magnetometers are a third kind of wearable sensor that is also explored in
activity recognition studies; their individual performance though is poor and they are
mostly used in combination with the other sensors. The aforementioned sensors are
found in all smartphones and smartwatches, which is the main reason they are widely
utilized for activity recognition studies since it is easy to extract their measurements.
The sensors are most often triaxial and they produce three vectors of raw signals,
one for each axis of the Cartesian reference system [93].
The raw data consist of three vectors of values, each one relevant to one axis
of the Cartesian system. After the extraction of the raw data, is the preprocessing
stage, which may include filtering and/or the normalization of the data to eliminate
signal noise. Features are afterwards extracted by a time window, which is employed
because it makes two signals comparable. Feature extraction retains valuable information from the signals [94]. The two basic categories of extracted features are time
domain and frequency domain and a list of the ones computed in most studies can
be found in [94]. After the extraction of features, a feature selection method may be
applied to identify which features will potentially assist in the improvement of the
recognition of activities and to eliminate large feature sets [94]. The HAR framework is concluded with the classification process, where a classification algorithm
is applied to recognize the activities.
Activity recognition tasks are actually multiclass classification problems. The
choice of the classification algorithm is driven by various parameters like the types of
A. Chatzimichail et al.
and meaning to the content of data coming from IoT devices and interlink related
data.
3.1 Modelling Sensors
Sensors embedded in devices that are attached on the human body or sensors directly
placed on the human body, are called wearable sensors [89]. The existence of such
sensors in mobile and wearable devices has led to their extensive use in activity
recognition and fall detection tasks. Accelerometers and gyroscopes are the most
popular ones, with accelerometers being the most effective in recognising activities
when used individually. Gyroscopes are also quite popular, however, they are mostly
used in combination with the accelerometers. Accelerometers are known to perform
well in recognizing activities in general but they are more successful in activities
with repetitive movement [89], since they measure a moving object’s magnitude and
direction. They usually fail to recognize similar activities when used individually
[90], thus it is more effective to use them along with other inertial sensors to improve
the performance of a human activity recognition system. Gyroscopes perform well
in recognizing an object’s orientation because they measure the rotation speed [91].
Gyroscopes are widely used in activity recognition studies, but most of the times
not as the only sensor. Fusion of such sensors, whether performed before or after the
classification algorithm, is found to improve the recognition rates of a human activity
recognition system, since one sensor may capture movements not well detected by the
other [92]. Magnetometers are a third kind of wearable sensor that is also explored in
activity recognition studies; their individual performance though is poor and they are
mostly used in combination with the other sensors. The aforementioned sensors are
found in all smartphones and smartwatches, which is the main reason they are widely
utilized for activity recognition studies since it is easy to extract their measurements.
The sensors are most often triaxial and they produce three vectors of raw signals,
one for each axis of the Cartesian reference system [93].
The raw data consist of three vectors of values, each one relevant to one axis
of the Cartesian system. After the extraction of the raw data, is the preprocessing
stage, which may include filtering and/or the normalization of the data to eliminate
signal noise. Features are afterwards extracted by a time window, which is employed
because it makes two signals comparable. Feature extraction retains valuable information from the signals [94]. The two basic categories of extracted features are time
domain and frequency domain and a list of the ones computed in most studies can
be found in [94]. After the extraction of features, a feature selection method may be
applied to identify which features will potentially assist in the improvement of the
recognition of activities and to eliminate large feature sets [94]. The HAR framework is concluded with the classification process, where a classification algorithm
is applied to recognize the activities.
Activity recognition tasks are actually multiclass classification problems. The
choice of the classification algorithm is driven by various parameters like the types of