On the Use of Decision Tree for Posture Recognition

Nooritawati Md Tahir Faculty of Electrical Engineering,

Universiti Teknologi Mara (UiTM),

40450 Shah Alam Selangor DE Malaysia.

norita_tahir@yahoo.com

Aini Hussain, Salina Abdul Samad,

Hafizah Hussin Dept of Electrical Electronics and System,

Faculty of Engineering and Building Environment,

Universiti Kebangsaan Malaysia

43600 Bangi, Selangor DE, Malaysia.

Abstract-The aim of this study is to evaluate the

effectiveness of decision tree as classifier for recognition of four

main human postures (standing, sitting, bending and lying)

since decision trees are well known for their success for

prediction, recognition and classification task in data mining

problems. Firstly, the eigenfeatures of these postures are

optimized via Principal Component Analysis rules of thumb

specifically the KG-rule, Cumulative Variance and the Scree

Test. Next, these eigenfeatures are statistically analyzed prior

to classification. In doing so, the most relevant eigenfeatures

that we termed as eigenpostures can be ascertained. Further,

we employed decision tree as classifier for posture recognition.

Initial results of the experiments are encouraging which

suggested that our method can efficiently be applied for

posture classification using DT.

Keywords-Decision Tree, Posture Recognition, Principal

Component Analysis, ANOVA

I. INTRODUCTION

As we know, pattern recognition is an integral part in

most machine intelligence systems built for decision making

tasks such as computer vision, biometric, document

classification and bioinformatics. The process of pattern

recognition can be divided into three principal steps namely

data acquisition, feature extraction and classification.

Classification is a major role in machine learning as well as

knowledge-based system, is the task of assigning objects to

one of several predefined categories or classes and Decision

Tree (DT) is among the classification technique that is

widely used. DT is a classifier in the form of tree structure

where each node is either a leaf node to indicate the value of

the target attribute or category of examples or a decision

node to specify some test to be carried out on a single

attribute-value with one branch and sub tree for each

possible outcome of the test. Selection of which attribute

should be primarily tested is based on the highest

information gain attained followed by recursive procedure.

The attractiveness of DT is it represents ‘rules’. Rules can

readily be expressed for ease understanding of human.

Some related work based on DT is by Yuan et al. [1] that

evaluated the performance of the classification of gene

sequence data using DT and SVM. Experiments conducted

proven that DT performed well as or close to SVM and even

better in some cases. Next, Lobato et al [2] have proved the

great versatility of DTs for real world applications

specifically to predict stochastic residual demand curves in

the Spanish electricity market, to estimate the daily load

pattern of units and to predict the values of reactors and

capacitors of the Spanish power system. Further, Li et al. [3]

utilized the heuristic information of decision tree for

improving its generalization capability. Conversely, Tsang

et al. [4] extended the model of DT classification to

accommodate data tuples having numerical attributes with

uncertainty described by arbitrary pdf’s. Results attained

showed that exploiting data uncertainty leads to DTs with

remarkably higher accuracies. Besides that, Kołakowska [5]

experimented DT for recognition of printed musical

symbols. Initial findings showed that recognition based on

DTs obtained higher accuracy than those based on

morphology and template matching implemented in Guido

system. Therefore, in this study we deemed further to

evaluate DT based on the algorithm known as Classification

and Regression Trees (CART) by Breiman et al. [6][7] for

posture recognition.

The structure of this paper is as follows. Section II

discusses the methodology, section III consists of results

and discussion and finally in section IV concludes our

findings.

II. METHODOLOGY

A. Overview of System.

An overview of the overall system that outlines the basic

structure is as depicted in Figure 1. It consists of the

following steps; pre-processing, feature extraction, feature

selection based on the rules of thumbs of Principal

Component Analysis (PCA) followed by statistical analysis

via ANOVA prior to classification. In the pre-processing

stage, silhouette of the subject is extracted based on

background subtraction and thresholding. Background

subtraction of image is differentiation between the present

frame image which contains the subject of interest and the

reference images. In this study, the background subtraction

technique used is as mentioned in [12]. Firstly, based on

motion of object, silhouette will be extracted from each of

the video sequence images. Next, morphological operations

namely dilation and erosion are employed. Both of these

2010 International Conference on Intelligent Systems, Modelling and Simulation

978-0-7695-3973-7/10 $26.00 © 2010 IEEE

DOI 10.1109/ISMS.2010.47

209

2010 International Conference on Intelligent Systems, Modelling and Simulation

978-0-7695-3973-7/10 $26.00 © 2010 IEEE

DOI 10.1109/ISMS.2010.47

209

operations can be manipulated using single or combination

of specific structuring elements. To eliminate all unwanted

noise, median filtering and morphological operations are

utilized. The main purpose of this step is to filter the

presence of noisy pixels in the foreground image. Further,

based on PCA and the three rules of thumbs of PCA, the

feature extraction phase is employed. It is a well-known fact

that the major goal for using PCA is to replace the p-

dimensional feature space with a much smaller m-

dimensional feature space, which nevertheless discards little

information.

For most empirical data, a large part of the total

variance can be sufficiently approximated with the first few

principal components only. However, the actual number of

principal components needed remains obscure. Therefore,

the three rules of thumb have been proposed for feature

selection as listed below:

i) Kaiser Gutman (KG) rule - The KG rule states that

any PC with a variance of less than one contain less

information than the original variables and is

therefore not worth retaining. In other words, the

KG-rule retains only those PCs whose variances, i.e.

eigenvalues that are ≥ 1. Nevertheless, for large

variable spaces p, the KG-rule usually retains too

many PCs [8] [9].

ii) Cumulative Variance - The criterion for choosing m

is to select a cumulative variance threshold, t where t

is at certain percentage of the total variance that the

first m PCs should account for. The required number

of PCs is then the smallest value of m for which the

chosen percentage is exceeded [9]. From PCA theory,

the variance of the i-th PC (eigenvector) is equal to

its corresponding eigenvalues λi. The accepted range

of percentage is within 80% and above [9].

iii) Scree Test involves looking at the plot of the

eigenvalues λi against the factor number k. The Scree

Test involves a certain degree of subjectivity since

there is no formal numerical cut-off based on the λi. If

the eigenvalues are plotted, they form a curve

heading towards almost 0% variance explained by the

last dimension. Thus, the point at which the curve

levels-out, sometimes referred to as the ‘elbow’

indicates the number of useful PCs, which are present

in the data [9].

Based on these three rules of thumbs, the significant

features which known as 'eigenpostures' are the eigenvectors

(principal components) of the set of posture images. Detail

description of the eigenpostures approach can be found in

[10] [11]. These eigenpostures will undergo the first stage of

feature selection process according to the three rules of

thumb mention above. Next, we will deem further by

applying ANOVA to these eigenpostures. As we know,

ANOVA is a technique for measuring the statistical

significance of a set of independent variables. The measure

that ANOVA produces is the p-value for the feature set. In

doing so, the groups that differ significantly are revealed. In

doing so, the optimized number of eigenpostures will be

determined for classification of the four main postures. Both

categories of eigenpostures will be classified using ANN

and DT.

B. Decision Tree (DT) Classifier

As mentioned, DT is a classifier in the form of a tree

structure, where each node is either a leaf node or a decision

node. The former indicates a class of instances while the

Figure 1: Overview of the overall system

Outcome:

Lying/Bending/Sitting/Standing

Raw Image

• Background Sub

• Thresholding &

Morph operations

Pre processing

• ANN & DT

Classifier

Classification

Feature Extraction

Generation of eigenpostures

based on PCA

• Phase I – 3 Rules of

Thumb of PCA

• Phase II – Statistical

Analysis

Feature Selection

210210

latter specifies some test to be carried out on a single

attribute-value, with one branch and sub-tree for each

possible outcome of the test. A DT can be used to classify

an example by starting at the root of the tree and moving

through it until a leaf node, which provides the classification

of the instance. CART is used in this study. CART generally

attempts to predict the values of a continuous variable from

one or more continuous and/or categorical predictor

variables. The splitting process is performed by choosing the

threshold value. It minimizes the impurity measure and used

as a splitting criterion. The CART algorithm uses the Gini

index to measure the class diversity in the nodes of the DT [6]

and produces four categorical outcomes. To be specific, the

categories of postures are standing, sitting, bending and lying

position. The effectiveness of DT in our classification task

will be evaluated.

C. ANN Classifier

ANN is a popular heuristic technique that can deal with

complex non linear problem even if the problem is

extremely complex to be translated in analytical form. It

deals with the training and testing processes before a

network can be precisely developed to perform the desired

task. The most exhaustive task in ANN is the training

process that requires numerous training patterns with

informative features or variables. Hence, feature extraction

and selection can be utilized to attain the most informative

variables that will speed up the convergence process.

III. EXPERIMENTS AND RESULTS

In this study, the aim is to evaluate and confirm the

capability of DT to classify eigenpostures that have

experienced the feature extraction and feature selection

phases. The generated eigenpostures are based on the

database of 800 images of various human postures that

include both standing and non-standing positions, bending

and lying with the human subjects are either facing front or

side with no restriction impose on the type of clothing being

worn. Firstly, each image has m x n pixels, and then

reshaped to a column vector of 1 x mn. Next, the

eigenvectors and eigenvalues are computed according to

Tahir et al. [11]. Further, the three rules of thumb of PCA

are implemented and hence the most suitable eigenpostures

can be ascertained.

Table 1 tabulated the results upon implementation of

KG rule, Cummulative Variance and Scree Test

respectively. Firstly, applying KG rule suggested that

retaining all eigenvalues > 1 results in thirty-five PCs to be

considered as significant components. Next, the cumulative

variance rule of thumb acted as feature selection to

determine the optimum number of eigenpostures or PCs.

From Table 1, the overall cumulative variance of the

eigenpostures is shown. As suggested in [9], a threshold t of

between 80%-90% can be considered to determine factor

n

TABLE 1

The Significant Eigenpostures Using The KG Rule,

Cummulative Variance And Scree Test

Factor k

KG Rule

Cumulative

Variance

Scree

Test

1 44.37 19.059 19.059

2 23.64 29.289 10.229

3 17.48 36.792 7.5032

4 12.25 42.074 5.2822

5 10.32 46.591 4.5166

6 9.37 50.671 4.08

7 5.77 53.214 2.5429

8 5.64 55.688 2.4741

9 5.21 58.015 2.3272

10 3.89 59.713 1.6985

11 3.79 61.388 1.6749

12 3.17 62.814 1.4258

13 3.07 64.177 1.3629

14 2.90 65.465 1.2879

15 2.64 66.673 1.2079

16 2.60 67.857 1.1841

17 2.48 69.01 1.1528

18 2.39 70.105 1.0947

19 2.19 71.099 0.99471

20 2.02 72.045 0.94541

21 1.97 72.949 0.90395

22 1.77 73.772 0.82319

23 1.71 74.571 0.79945

24 1.67 75.312 0.74046

25 1.63 76.038 0.72617

26 1.57 76.73 0.69188

27 1.42 77.358 0.62865

28 1.35 77.977 0.6185

29 1.28 78.551 0.57428

30 1.26 79.121 0.56953

31 1.20 79.676 0.55543

32 1.14 79.702 0.52608

33 1.11 79.79 0.52042

34 1.06 80.21 0.52036

35 1.01 81.678 0.52025

36 0.984 82.142 0.5017

37 0.982 82.593 0.45103

38 0.981 83.02 0.42684

39 0.981 83.436 0.41549

40 0.980 83.837 0.40115

Note: Italic Values Indicate the optimized eigenpostures

subsequent to statistical analysis

211211

number, k. In this case, an 80% criterion would result in k

equals 34 as tabulated in Table 1. Finally, the Scree test

outcome in Table 1 illustrated the decrease in magnitude for

successive eigenvalues implies that the first few principal

components can approximate a large part of the original

data’s variance. In this case, decision to retain the first

thirty-five PCs is appropriate and they reasonably well to

represent good approximation of the original data set. These

eigenpostures are known as Type I and further undergo the

statistical analysis prior to classification. Accordingly, the

statistical significance of all Type I eigenpostures of the four

main postures is determined via ANOVA. Hence, from the

ANOVA test, at a significant level of α = 0.05, the outcome

for p-values of eigenpostures 1-9, 11-13, 15-18, and 20-22

are numerically indistinguishable from zero. As a result, the

ANOVA test has lucratively reduced the feature vectors to

nineteen or 54% of the initial feature extraction quantity and

these eigenpostures are known as Type II eigenpostures.

As aforementioned, ANN along with DT was chosen as

classifiers in this study. A three-layer NN with weights

adjusted using the Levenberg-Marquardt was trained to

determine the relationship between the selected

eigenpostures and the respective four posture classes namely

standing (ST), sitting (SI), bending (BD) and lying (LY).

Table 2 tabulated the classification result of Type I and

Type II with ANN and DT as classifiers. It is observed that

lying postures gained 100% accuracy rate for both

categories of eigenpostures for ANN as well as DT. This is

due to the nature of lying position that is extremely distinct

as compared to the other three postures.

Further, for the bending posture, the recognition rate

attained by Type II eigenpostures is higher than Type I

specifically 98% as compared to 94% for ANN. As for the

DT, both type of eigenpostures attained equal recognition

rate specifically 98%. Further, for sitting posture, both

categories achieved equal recognition rate that is 98% for

ANN but for DT, Type II eigenpostures obtained better

accuracy. For the standing posture, both classifiers attained

equal classification rate namely 98%. On the whole, for

ANN as classifier, Type II eigenpostures obtained average

recognition rate of 98.5% for ANN as well as DT which is

higher than Type I eigenpostures. It is proven that DT is

capable to perform posture recognition effectively as ANN.

Figure 2 displays the CART generated based on both

Type I and II eigenpostures with 10 maximum tree depth for

classifying the four possible outcome that is standing (ST),

sitting (SI)’, bending (BD)’ and lying (LY). The CART

algorithm selects E1, E2, E3, E4, E5, E6 and E11 as the seven

main attributes and picks E1 as the top node in the

discrimination process. Beginning from the top node, the

rule of ‘E1 < 0.014’ classifies the test image attribute

accordingly into any one of the fifteen possible leaf nodes.

If the top rule is satisfied, the decision takes the left path or

vice versa. The posture category that is bracketed in italic

form ‘(ST)’; is the recognition category if that specific level

tree depth is selected. Ultimately, a decision is reached

when a leaf node assigned the test image or observation as

either ‘ST’, ‘SI’, ’BD’, or ’ LY’. One interesting fact noted

is that from Figure 2, the DT selected the seven top

attributes for construction of the tree prior to statistical

analysis.

TABLE 2

CONFUSION MATRIX FOR POSTURE RECOGNITION BASED ON ANN and DT

PREDICTED CATEGORY

Type I Eigenpostures

(35 eigenpostures)

Type II Eigenpostures

(19 eigenpostures)

ANN DT ANN DT

ACTUAL

CATEGORY

BD SI ST LY BD SI ST LY BD SI ST LY BD SI ST LY

BD 188 4 8 0 196 2 2 0 196 0 4 0 196 0 4 0

SI 0 196 4 0 2 194 4 0 4 196 0 0 4 196 0 0

ST 4 0 196 0 2 2 196 0 4 0 196 0 2 2 196 0

LY 0 0 0 200 0 0 0 200 0 0 0 200 0 0 0 200

Accuracy (%) 97.5 98.25 98.5 98.5

212212

Figure 2. The Decision Tree Generated for Posture Classification for Type I and Type II Eigenpostures

G

M

N

F E

K L

I

D

A

B H

C J

LY ST

LY ST ST ST

ST

ST

SI ST

ST

SI

ST

SI ST

Node A: E1 < 0.014 (ST)

Node B: E1 < - 0.273 (LY)

Node C: E2 < 0.233(LY)

Node D: E3 < -0.151 (ST)

Node E: E4 < 0.164(LY)

Node F: E2 < -0.178 (ST)

Node G: E1 < -0.11 (ST)

Node H: E4 < 0.122 (SI)

Node I: E3 < 0.097 (SI)

Node J: E11 < -0.086 (ST)

Node K: E6 < -0.242 (SI)

Node L: E5 < - 0.069 (ST)

Node M: E1 < 0.13(ST)

Node N: E < 0.155 (SI)

Type II

Type I

1 2 3 4 5 6 7 8 9 100

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Number of Tree Depths

Cla

ss

ific

ati

on

Ra

te (

x1

00

%)

Figure 3. The accuracy rates with different number of tree depths. The best performance is at maximum tree depth.

213213

Additionally, Figure 3 shows the relationship between the

classification rate and the number of tree depth used. The

best classification rate attained for Type I is 98.25% while

for Type II is 98.5% when all tree depths were utilized. At

level seven tree depth, the classification accuracies attained

reduced below 90% for both type of eigenpostures.

IV. CONCLUSION

In conclusion, this study has proven the ability of DT to

perform recognition of four main human postures as good as

ANN. From the constructed DT, the seven feature vectors of

eigenpostures are confirmed as the most suitable and

optimized in discriminating human postures with above

98% accuracy rate for both categories. This work could be

extended to a broad range of posture recognition for safety

measures, intruder’s alertness, and action recognition for

surveillance applications.

REFERENCES

[1] X. Yuan, X. Yuan, F. Yang, J. Peng and B.P. Buckles, “Gene

Expression Classification: Decision Trees vs. SVMs,” Proceedings of the Sixteenth International Florida Artificial Intelligence Research Society Conference, 2003, USA, pp 92-97.

[2] E. Lobato, A. Ugedo, L. Rouco, and F.M. Echavarren, “Decision Trees Applied to Spanish Power Systems Applications,” Proceedings

of 9th International Conference on Probabilistic Methods Applied to Power Systems KTH, Stockholm, Sweden - June 11-15, 2006.

[3] N. Li, L. Zhao, Ai-Xia Chen, Qing-Wu Meng and Guo-Fang Zhang, “A New Heuristic Of The Decision Tree Induction,” Proceedings Of The Eighth International Conference On Machine Learning And Cybernetics, Baoding, 12-15 July 2009.

[4] S. Tsang, B.Kao, Kevin Y. Yip, Wai-Shing Ho and S,D. Lee, “Decision Trees for Uncertain Data”, Proceedings of IEEE International Conference on Data Engineering, 2007.

[5] Agata Kołakowska, “Applying Decision Trees to the Recognition of Musical Symbols,” Proceedings of the 2008 1st International Conference on Information Technology, IT 2008, Gdansk, Poland, 2008.

[6] Breiman L., Friedman J.H., Olshen R.A. and Stone C.J., “Classification and Regression Tree,” Belmont, CA. Wadsworth International Group, 1984.

[7] Tom Mitchell, “Machine Learning”, McGraw Hill, 1997.

[8] Nooritawati Md Tahir, Aini Hussain, Salina Abdul Samad, Khairul Anuar Ishak and Rosmawati Abdul Halim. “Feature Selection for Classification Using Decision Tree” Proceeding of SCORed 2006, Malaysia. 2006, pp 99-102.

[9] IT Jolliffe. Principal Component Analysis, Springer Series, 2001.

[10] M Turk and A. Pentland. “Eigenfaces for Recognition”. Journal of Cognitive Neuroscience. 3(1): 71-86, 1991.

[11] Nooritawati Md Tahir, Aini Hussain, Salina Abdul Samad, Hafizah Husain and Mohd Marzuki Mustafa,“Eigenposture for Classification” Journal of Applied Sciences 6(2): 419-424, 2006.

[12] S. Sulaiman, A.Hussain, N. Md Tahir, S. A. Samad. ”Graphical User Interface (GUI) Development for Object Tracking System in Video Sequences,” World Applied Sciences Journal, 4(2), pp 244-259.

214214