IEEE Robotics & Automation Magazine - September 2015 - 119

human motion primitives
(HMP); the right column of
Table 2 lists some of the HMP
associated with the locationindependent ADL.
Most wearable sensing systems devoted to the recognition of HMP rely exclusively
on acceleration information
and share a similar architecture [3], graphically summarized in Figure 3. The main
tasks of the system are: 1) to
extract relevant features from
the available sensory data, 2)
to create representations of
the target HMP in terms of
the available features, and 3)
to classify the run-time sensory data according to the
known representations.

ADL

Cognitive ADL
Responsibility to Handle
Own Finances

Mechanical ADL
Feeding

Involuntary ADL
Continence

Future
Research

Ad Hoc
Solutions

Outdoor ADL
Shopping

In-Home ADL
Housekeeping

Location-Independent ADL
Transferring

Smartphones

Smart Environments

Wearable Devices

WHARF Data Set
The WHARF data set is composed of over 1,000 acceleraFigure 2. By considering the cognitive abilities and environment characteristics required for the
tion recordings of all the successful execution of an ADL it is possible to identify its most suitable monitoring technique.
HMP listed in Table 2, col- The state-of-the-art monitoring systems allow for the recognition of: outdoor ADL via smartphonelected from 17 volunteers (11 based systems, in-home ADL with smart environments, and location-independent ADL with
wearable sensing systems.
men and six women, ages 19
to 81). The data set allows for
the creation and the validation of the HMP models. By anaTable 2. The location-independent ADL and HMP.
lyzing the performance of different modeling and classification algorithms over the same validation data set it is possible
ADL
Motion Primitive
to compare different solutions.
Toileting
Brush own teeth
Each recording reports the triaxial acceleration values regComb own hair
istered during one execution of one motion by an ad hoc
Transferring
Get up from the bed
sensing device worn at the right wrist. As summarized in
Lie down on the bed
Table 1, the device, shown in Figure 4, is equipped with a sinSit down on a chair
Stand up from a chair
gle triaxial accelerometer with measurement range of
[- 1.5 g; + 1.5 g] and sensitivity of 6 b/axis, working at the
Feeding
Drink from a glass
Eat with fork and knife
sampling frequency of 32 Hz.
Eat with spoon
All the trials were recorded at the homes of the volunPour water into a glass
teers in a series of supervised experiments. During each exAbility to use telephone Use the telephone
periment, a volunteer equipped with the device is asked to
Indoor transportation
Climb the stairs
perform one of the motions of interest and the supervisor
Descend the stairs
labels the corresponding acceleration data as an execution
Walk
of the motion. To increase the naturalness of the executions
in each session, the supervisor defines which motions are
to be performed and the number of repetitions, which var- ● the creation of HMP models (models builder)
ies with the motion and the volunteer, while the volunteer ● the classification of unknown recordings on the basis of
is allowed to choose which motion to perform each time.
existing models (classifier).
The models builder library [5] executes the steps, shown
HMP Detection Library
in Figure 5, to create the model of an HMP of interest on the
The HMP detection library is a public library of MATLAB basis of a representative modeling data set. The WHARF
functions designed to work with the WHARF data set and al- data set provides both specific, i.e., biased toward the behavlows for:
ior of one individual, and general, i.e., not biased, modeling
September 2015

*

IEEE ROBOTICS & AUTOMATION MAGAZINE

*

119
Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - September 2015
