IEEE Systems, Man and Cybernetics Magazine - October 2020 - 8

to detect eight sitting postures
met hod s. Ba sed on d i f ferent
with 95.5% accuracy.
classification approaches, these
There is increasing
Other studies exploited an 8 ×
studies mainly consist of machineinterest in the
8 sensor matrix, and machinelearning methods and thresholdlearning methods were used to
based algorithms.
development of
recognize the postures. Kamiya et
Works that adopted machinetechnologies that
al. [18] used an 8 × 8 pressure senlearning methods are discussed in
monitor and support
sor matrix, and a radial basis
this section. Multu et al. [22] used
function-based support vector
the Tekscan [13] to detect signifiseated users.
machine (SVM) algorithm was
cant pressure-sensing points, and
used to recognize postures. Fard
then, to support the recognition of
et al. [19] proposed a system using
sitting postured, it identified an
an 8 × 8 pressure sensor matrix on the seat to prevent
optimal deployment composed of 19 sensors. Hu et al. [23]
pressure ulcers. Kim et al. [20] developed a sensing cushproposed the use of PoSeat equipped with an acceleromeion by mounting 8 × 8 pressure sensors to classify five
ter and pressure sensors. A hybrid SVM classifier was used
kinds of children's sitting postures. The authors proto recognize sitting postures for chronic back pain prevenposed a CNN-based method and compared the results
tion. Benocci et al. [24] proposed a system based on five
with other approaches [naïve Bayes, Decision Tree (DT),
pressure sensors and k-nearest neighbor (kNN) to detect
a regression model, and the Support Vector Machine].
five postures. Min [25] developed a real-time sitting posCNN obtained the highest recognition results, with an
ture monitoring system (SPMS) that prompts users to keep
average accuracy of 95.3%. Xu et al. [21] presented a
correct postures. Zemp et al. [26] developed an instrumentcushion deployed on the seat and backrest, and the
ed chair with force and acceleration sensors, with five difpressure sensors data were converted to binary values
ferent machine-learning methods compared, resulting in
("true" or "false," according to a given pressure threshan average accuracy of 90.9%. Fu and Macleod [27] proold) to recognize nine postures. All of the analyzed literposed a system based on eight pressure sensors placed on
ature works using smart cushions with dense pressure
a chair backrest and seat. A hidden Markov model (HMM)
sensor arrays are summarized in Table 1.
was adopted to analyze sitting posture sequences. Kumar
et al. [28] designed a system named Care-Chair, using four
pressure sensors on the backrest to recognize users' comSmart Cushions With Sparse Pressure
plex sedentary and emotion-related activities.
Sensor Arrays
In our previous research [29], we used three pressure
Dense pressure sensor arrays are costly solutions, so sevsensors placed on a smart wheelchair seat to detect
eral studies proposed sparse pressure sensor array
users' sitting postures. Liang et al. [30] proposed a practical sitting posture recognition system using a sparse
Table 1. The state of the art on smart
pressure sensor array. User-invariant features were
cushions with dense pressure sensor
arrays.
extracted and the sitting posture was recognized using
an AdaBoost classifier. They also proposed two protoSensor Array
Placement of the
type applications of video game control and wheelchair
Reference
Type
Cushion
control using the sitting posture recognition results,
Xu et al. [12]
E-textile
Seat
including lean left (LL), lean right (LR), lean forward
(LF), and sit upright.
[13]
E-textile
Seat and backrest
Also, in [31], the authors demonstrated an effective
Tan et al. [14]
E-textile
Seat and backrest
sensor placement and an ensemble learning classifier
capable of recognizing 15 fine-grained postures with an
Mota and Picard [15]
E-textile
Seat and backrest
accuracy of 98%. Roh et al. [32] proposed an SPMS
Meyer et al. [16]
Textile pressure
Seat
using four low-cost load cells. Six postures could be
sensor
recognized and a 97.2% classification accuracy was
Liu et al. [17]
32 × 32 sensor
Seat
achieved using an SVM classifier. Bibbo et al. [33]
array
adopted pressure cushions on the backrest and seat to
assess cognitive engagement based on sitting recogniKamiya et al. [18]
8 × 8 sensor array
Seat
tion results. Ren et al. [34] suggested a health-promotXu et al. [21]
Seat 6 × 8, back- Seat and backrest
ing system for relaxation and fitness microbreaks at
rest 2 × 8
work. In addition to cardiac monitoring, the authors
Fard et al. [19]
8 × 8 sensor array
Seat
detected sitting behaviors using an artificial neural network applied to data acquired from six pressure senKim et al. [20]
8 × 8 sensor array
Seat
sors placed on the seat.
8

IEEE SYSTEMS, MAN, & CYBERNETICS MAGAZINE O ctober 2020
IEEE Systems, Man and Cybernetics Magazine - October 2020

Table of Contents for the Digital Edition of IEEE Systems, Man and Cybernetics Magazine - October 2020
