1 - Human Sitting Postures Classification Based On Angular Features With Fuzzy-Logic Labeling
1 - Human Sitting Postures Classification Based On Angular Features With Fuzzy-Logic Labeling
Abstract – Musculoskeletal disorders (MSDs) are generally associated with sitting postures. Assessing
and ensuring healthy sitting posture are indispensable aspects of reducing the occurrence of MSDs. This
study aims to develop a system that allows office workers' body postures to be contactless and recognized
by different classification methods while sitting on a chair and can be used for health applications. Five
different sitting body postures have been determined within the scope of medical and health literature
studies and relevant standards. Thirty subjects were asked to sit in these body postures for 30 seconds.
While the subjects were sitting, skeleton point position defined as a pose data of the subjects were
obtained from the Kinect device simultaneously. Five angles that are thought to distinguish sitting
positions according to different joint positions were determined and calculated. The angle values that can
take in the standard sitting position in the literature have been determined. According to these values, the
angle values in other postures were determined. A rule-based fuzzy inference system was designed using
angle values for labeling sitting posture data. Angle values were calculated to classify the labeled depth
values, and an artificial neural network classifier was designed according to these angle values. As a
result, five different sitting body postures were classified with KNN (K-Nearest Neighbours) and Neural
Network (NN), respectively, with 98.9% and 97% overall accuracy values. The study was compared with
other studies in the literature. In this context, a high-performance system design that can improve healthy
sitting behaviors for office workers that can be used in both health applications and robot vision is
presented.
Keywords – Sitting Posture, Angular Features, Fuzzy-Logic Labeling, Classification, Depth-Based Sensor, Kinect
615
has been limited to 0.001 [23], [24]. The weights
and biases are initialized using the Nguyen-
Widrow method. For the KNN, model flexibility
parameters such as the number of neighbors,
distance metric, and distance weight have been
chosen as 3, Euclidean, and Uniform, respectively.
All models were validated through a 5-fold cross-
validation (cv) process to evaluate the predictive
ability as it allows the classifier to operate without Figure 8 Performance indicators of all classifiers
bias and avoids the overfitting problem. The cv Finally, it is seen that the samples labeled as 5th
was performed without data sharing between class are mainly classified as 1st and 4th class,
training and validation data to avoid overtraining. except for their groups. When the performance of
In order to measure the performances of each the classifiers is evaluated, if Figure 7 and Figure
model, a multi-class confusion matrix which is 8 are examined, it is seen that the classification
defined in [25] and the ROC curve, is created, and accuracy of the KNN classifier is higher than the
Accuracy (A), Recall (R), precision (P), F1-score NN classifier.
(F), AUC (Area Under Curve), LogLoss (LL-
logistic loss) and Specificity (S) indicators are
calculated to evaluate performance [25].
IV. RESULTS AND DISCUSSION
For the classification of body postures,
confusion matrices for the models of NN, KNN,
and classifier methods with the highest accuracy
are presented in Figure 7. Both training and testing
processes were performed on the same computer.
Figure 9 ROC curves of classifiers
When the confusion matrices are examined, it is
seen that the samples are mainly classified
The ROC curves belonging to the models with
according to their classes.
the highest accuracy values for interpreting the
accuracy values are presented in Figure 9. When
the ROC graphs are examined, it is seen that all
models are very close to the upper left corner point
(0,1); therefore, the ability of the models to
diagnose classes fits quite well. Different studies
classify the data from different skeleton points in
the related literature. However, in these studies,
data acquired from the specified skeleton points
were classified at once, and the data from different
Figure 7 Confusion matrices of classifiers
skeleton points were not separated and classified
It is seen that the first-class labeled samples are
same time. Therefore, when True Positive Rates
mainly classified as 5th class apart from their
(TPR) in confusion matrices and Table III are
groups and vice versa. Next, it is seen that the
examined, the best and worst overall accuracy
samples labeled as 3rd class are mainly classified
classes are the second and fifth classes.
as 4th class and 2nd, except for their groups, and
the samples labeled as 4th class are mainly
classified as 3rd class, except for their groups.
616
Table III. Performances of classification methods for all this context, classification accuracy values and
posture classes other features obtained in studies in the relevant
literature were compared with the results of this
Posture
A F1 P R
Model
Time-based
Fuzzy rule
calculation
calculation
calculation
Automatic
Geometric
Labeling
calculator
Angular
Angular
Angular
Method
KNN
shape
set
4 0.991 0.991 0.961 0.960 0.962 0.192 0.989
Statistics features
physical features
Grayscale image
Angular feature
5 0.984 0.988 0.934 0.938 0.930 0.285 0.984
Local contour -
Human body
Feature
topological
Method
Features
Angular
In order to visualize the sitting posture data, the
data in each sitting class and sample 3D drawings
of them are given in Figure 10. When these City-
BP
Classifier Block LDA KNN NB KNN
drawings are examined, it is seen that the TP, FP, Distance
NN
and FN groups differ.
Joint
16 12 6 10 8 6
Point
The
Number 2 4 9 2 2 5
of Classes
Pose
5600 22226 16200 397800 55080 49580
Number
The
Number
20 8 18 28 100 30
of
Subjects
618
Engineering Informatics, vol. 26, no. 2, pp. 439–455, Apr.
2012, doi: 10.1016/J.AEI.2012.02.011.
[12] N. J. Delleman and J. Dul, "International standards on
working postures and movements ISO 11226 and EN
1005-4," https://doi.org/10.1080/00140130701674430,
vol. 50, no. 11, pp. 1809–1819, Nov. 2007, doi:
10.1080/00140130701674430.
[13] J. Kelly, "Proper Height For Standing Desks," 2014.
https://notsitting.com/proper-height/
[14] J. Elliott, "How long should I stand at my standing desk? -
HealthPostures," 2020. https://healthpostures.com/how-
long-should-i-stand-at-my-standing-desk/ (accessed Jun.
23, 2022).
[15] Fondazione Ergo-MTM Italia, “Ergonomic Assessment
Worksheet,” 2021. [Online]. Available:
https://www.eaws.it/
[16] Canadian Centre for Occupational Health & Safety,
"Working in a Sitting Position -Good Body Position,"
2017.
[17] "ISO - ISO 7250-1:2017 - Basic human body
measurements for technological design — Part 1: Body
measurement definitions and landmarks."
https://www.iso.org/standard/65246.html (accessed Dec.
24, 2022).
[18] Y.-K. Kong, S. Lee, K.-S. Lee, and D.-M. Kim,
"Comparisons of ergonomic evaluation tools (ALLA,
RULA, REBA and OWAS) for farm work," International
journal of occupational safety and ergonomics, vol. 24,
no. 2, pp. 218–223, 2018.
[19] J. Vince, "Mathematics for Computer Graphics," 2022,
[20] J. Liu and Y. Zhang, "An Attribute-Weighted Bayes
Classifier Based on Asymmetric Correlation Coefficient,"
Intern J Pattern Recognit Artif Intell, vol. 34, no. 10, Jan.
2020, doi: 10.1142/S0218001420500251.
[21] R. K. Bania, "R-GEFS: Condorcet Rank Aggregation
with Graph Theoretic Ensemble Feature Selection
Algorithm for Classification," Intern J Pattern Recognit
Artif Intell, vol. 36, no. 9, Jun. 2022, doi:
10.1142/S021800142250032X.
[22] T. Hu et al., "Rice Variety Identification Based on the
Leaf Hyperspectral Feature via LPP-SVM," Intern J
Pattern Recognit Artif Intell, vol. 36, no. 15, p. 2350001,
Jan. 2023, doi: 10.1142/S0218001423500015.
[23] H. Coskun and T. Yigit, "Artificial Intelligence
Applications on Classification of Heart Sounds," in
Nature-Inspired Intelligent Techniques for Solving
Biomedical Engineering Problems, IGI Global, 2018, pp.
146–183.
[24] H. Coskun, O. Deperlioglu, and T. Yigit, “Ekstra Sistol
Kalp Seslerinin MFKK Öznitelikleriyle Yapay Sinir
Aǧlari Kullanilarak Siniflandirilmasi,” 2017 25th Signal
Processing and Communications Applications
Conference, SIU 2017, Jun. 2017, doi:
10.1109/SIU.2017.7960252.
[25] H. Coskun, T. Yiğit, İ. S. Üncü, M. Ersoy, and A. Topal,
"An Industrial Application Towards Classification and
Optimization of Multi-Class Tile Surface Defects Based
on Geometric and Wavelet Features," Traitement du
Signal, vol. 39, no. 6, pp. 2011–2022, Dec. 2022
[26] F. Bozkurt, H. Coskun, and H. Aydogan, "Effectiveness
of Classroom Lighting Colors Toward Students' Attention
and Meditation Extracted From Brainwaves," JOURNAL
OF EDUCATIONAL AND INSTRUCTIONAL
STUDIES, vol. 4, no. 2, pp. 6–12, 2014.
[27] H. Aydogan, F. Bozkurt, and H. Coskun, "An assessment
of brain electrical activities of students toward teacher's
specific emotions," International Journal of Psychological
and Behavioral Sciences, vol. 9, no. 6, pp. 2037–2040,
2015.
619