State of the Art - Smartphone-Based Human Activity Recognition - page 48

them are: (Zhao et al. 2010 ), a cross-people AR technique is described. They transfer
classification knowledge to a new user without data collection by adapting the activ-
ity labeled samples of an initial model using a binary DT model. Also in Allen et al.
( 2006 ), a method for user adaptation based on GMM is proposed to compensate the
problem of limited training data for each test subject.
3.3.4 Offline and Online
HAR system can also be classified into two main groups depending on the response
time that systems take to perform activity classification (Lara and Labrador 2012a ):
online methods aim for real-time prediction of activities while, conversely, offline
methods usually need extra processing time, use computationally demanding algo-
rithms or simply they do not require real-time operation. In the first case, many of
these approaches focus on the detection of a small group activities (generally SPs
and AAs) because an increment in the number of activities or their complexity might
not allow to achieve real-time operation.
Various previously proposed systems have required external computing support
for allowing online recognition capability. This has been done by transmitting body
motion data in real-time to a fixed device. For example, a real-time system for the
detection of basic SPs and AAs was described in Karantonis et al. ( 2006 ). It used
a waist-mounted wireless unit composed of a microcontroller and an accelerometer
which transmitted processed inertial signals to a local computer for activity evalu-
ation and display. Furthermore, an online approach was introduced in Tapia et al.
( 2007 ). It classified 30 physical gymnasium activities and their intensities using 3
accelerometers and a heart rate monitor. Signals were also wirelessly transmitted to
a laptop for processing and classification. In Ermes et al. ( 2008 ), motion bands with
an embedded accelerometer attached to the user's wrist, ankle and chest transmitted
via Bluetooth link the inertial signals to a PDA for the classification of fitness-related
activities. Nowadays PDAs are nearly obsolete and are being replaced with smart-
phones (Liu et al. 2011 ).
Smartphones have simplified the online implementation of many applications as
they integrate sensing and computing capabilities, in contrast to the aforementioned
approaches which had to rely on ad-hoc implementations to establish the recogni-
tion pipeline using different devices. First smartphone-based HAR attempts were
offline and only recorded the inertial data using the device sensors or other linked
devices and later processed the signals/data for activity classification. For example: in
Lara et al. ( 2012 )the Centinela system was presented. It consisted of a chest unit
composed of several sensors to measure acceleration data and vital signs (e.g. heart
rate, breath amplitude, respiration rate) and a smartphone wirelessly connected via
Bluetooth. Data was later processed and classified offline using different ML algo-
rithms showing how vital signs can also add value to the system performance.
Kwapisz et al. ( 2011 ), which was previously mentioned, also developed an offline
smartphone-based HAR system for the classification of 2 SPs and 4 AAs. In the