State of the Art - Smartphone-Based Human Activity Recognition - page 35

their daily life and delivers context-aware data that can help to provide valuable
services, products and technologies aiming to improve people's QoL.
In this chapter we address the subject of HAR supported by its most relevant works
in the literature. In Sect. 3.2 , we start with a general description of this field includ-
ing aspects such as the standard recognition process, studied activities, common
approaches for experimental data collection, and performance evaluation. Subse-
quently, we emphasize on the state of the art research on HAR systems, some of
which are described and compared against our approaches in Sect. 3.3 . Finally, we
conclude the state of the art in Sect. 3.4 .
3.2 Human Activity Recognition Overview
HAR is an essential component for the development of systems for allowing smarter
interactive cognitive environments. As a mode of illustration, in a simplified view of
the human information processing pipeline which consists of four stages ( sensing ,
data analysis , decision making and taking action ) (Gandetto et al. 2003 ; Parasuraman
et al. 2000 ), HAR belongs to the first and second stages. It contributes to acquire
the necessary information regarding the user activity. This is then combined with the
perceived environmental data in order to obtain a complete state representation of
the world and its individuals before continuing to subsequent stages. For instance, a
system for the management of accidents in the elderly (e.g. falls (Lord 2007 )) would
require initially the detection of a potential event from the user's wearable sensors.
Then this detection needs to be fused with environmental information which can
help to confirm that what has occurred is not a false positive (e.g. if the sensor felt
on the floor). Afterward, it is possible to make decisions about what to do and if
necessary take action by calling the emergency services for immediate assistance. In
this section we introduce the main concepts behind HAR and their application into
real world problems with particular focus on the fields of AAL and AmI.
Ageneral representation of the principal components of a HAR process is depicted
in Fig. 3.1 . Many of the HAR approaches found in literature, follow a regular struc-
ture with slight variations based on their application, sensors, and selected ML algo-
rithms. The diagram is valid to supervised, semi-supervised and incremental learning
approaches (Karantonis et al. 2006 ; Stikic et al. 2011 ; Wang et al. 2012 )(Referto
Sect. 2.5.1 ) . They differ on the type of input (labeled or unlabeled) and if the learned
model updates when new samples are added into the system (notice the Feedback
dotted line on the graph).
From the four main blocks of a HAR system (Fig. 3.1 ), sensing is responsible of
gathering the sensor data from the available sources and process them (see Sect. 2.3 ) .
Generally, signal conditioning (e.g. reducing noise, digitizing, amplifying) is always
required for adapting the sensed signals to the application requirements. In the second
place, the feature extraction process is in charge of obtaining meaningful features
that describe the data and allow a better representation and understanding of the
studied phenomena. The extracted features turn into the input of the ML algorithm,