State of the Art - Smartphone-Based Human Activity Recognition - page 36

Feedback

Learned
Model

Learning

Feature
Extraction

Sensing

Predicted
Activity

Prediction

Fig. 3.1 The human activity recognition process pipeline with its four main blocks

either for learning the model or for the activity prediction of novel samples when
the model already exist.
Moreover, traditional HAR systems usually operate in a feed-forward basis thus
learning is performed offline only once and there is no further feedback into the
system. This is useful in cases where the data distribution does not change over
time or the system is subject-independent and robust against high input variabil-
ity. Otherwise adaptive methods such as incremental online or transfer learning
(Zheng et al. 2009 ) are advised but conditioned with an increase in the computa-
tional load into the process. In relation to the analysis of high level activities, which
are combinations of simple activities (e.g. assembling furniture or fixing a car (Amft
et al. 2007 )), there is a limited amount of work that has been done and it is still an
open research field.

3.2.1 Human Activities

A HAR system is dependent on the set of activities to recognize as they can directly
affect the way systems are designed and implemented. For this reason, the classifi-
cation of activities in categories simplifies the selection of appropriate mechanisms
to recognize activities. Here, we categorize them with respect to their duration and
complexity, and activity type.
Table 3.1 shows the categorization of activities with respect to their duration and
complexity. There are three main groups: Short events are simple activities with a
defined small duration. These are divided in two types. First, Body gestures which are
visible motions mainly used as a mechanism for nonverbal communication and sec-
ond, transitions which are the events that connect the execution of two different activ-
ities. For example: when a person is seated and then stands up, the (SiSt) PT occurs.