3D Face Recognition on Point Cloud Data

An Approaching based on Curvature Map Projection using Low Resolution Devices

Luis Felipe de Melo Nunes

, Caue Zaghetto

and Flavio de Barros Vidal

Department of Computer Science, University of Brasilia, Brazil

Department of Mechanical Engineering, University of Brasilia, Brazil

Keywords:

Point Cloud, Face Recognition, Curvature Maps, Three-dimension Face Data, Low Resolution Device.

Abstract:

Facial recognition is the most natural and common form of biometrics, routinely used by humans and one of

the most promising areas in biometrics research. The majority of traditional researches and commercial use

of facial recognition systems are focused on methods that explores 2D (two-dimensional) images of human

faces. All of them are based on features extraction that does not use any 3D shape information from the

faces, especially with regard to depth. This paper presents a method based on Point Cloud and Curvature

Map Projection to perform a 3D face recognition. The achieved results are presented and divided in two

test scenarios, composed by a biometric evaluation analysis applying the Equal Error Rate score, Receiver

Operating Characteristic and an accuracy comparison with other related works. The proposed work presents

an accuracy of about 98.92%, allowing it to be applied for 3D face recognition tasks.

1 INTRODUCTION

The increasing need to monitor and restrict access

to information or environments has led to major ef-

forts towards the development of a variety of security

mechanisms, such as biometric systems (Jain et al.,

2000). In addition to applications related to access

control, there are also others associated with civil

identiﬁcation and criminal investigation. To properly

identify a user, biometric systems must rely on traits

that present sufﬁcient levels of universality, distinc-

tiveness, permanence, collectability, acceptability and

circumvention (Jain et al., 2004).

Among the various ways of performing biomet-

rics, it is possible to highlight the facial recognition.

Undoubtedly, facial recognition is the most natural

and common form of biometric routinely used by hu-

mans and one of the most promising areas in biomet-

rics research (Soldera et al., 2017). In general, facial

recognition algorithms uses facial shape and their spa-

tial relationships to perform individuals recognition

(Jain et al., 2000). Although a human being is able to

recognize a human face in an unfamiliar environment

in approximately 100-200 ms, to a computer, running

the best yet existing algorithms, it is still a challenge

to accomplish this kind of task (Haykin and Network,

2004). It is true that in the last decade the reliability

of face recognition algorithms has been improved, but

in unconstrained environments problems such as un-

controlled illumination, head pose, facial expression

and partial occlusion are still a bottleneck to these al-

gorithms to achieve higher efﬁciency (Soldera et al.,

2017).

The majority of traditional research and commer-

cial use of facial recognition systems are focused on

methods that explores 2D (two-dimensional) images

of human faces (Bowyer et al., 2006). These meth-

ods, in general, are based on features extraction that

does not take into account the 3D shape of faces, es-

pecially with regard to depth.

This paper presents a method based on Point

Cloud and curvature map projection in order to per-

form face recognition. To the best of our knowledge,

although some works have already addressed the chal-

lenge of 3D facial recognition (Patil et al., 2015), the

solution presented in this paper is the only one that

uses Point Cloud data and FCM method, applied to

a dataset of 3D face images acquired by a low cost

sensor device, to perform the task.

The remainder of this paper is divided into ﬁve

sections: In Section 2, related works are discussed;

Background concepts and the proposed method are

266

Nunes, L., Zaghetto, C. and Vidal, F.

3D Face Recognition on Point Cloud Data - An Approaching based on Curvature Map Projection using Low Resolution Devices.

DOI: 10.5220/0006843702660273

In Proceedings of the 15th International Conference on Informatics in Control, Automation and Robotics (ICINCO 2018) - Volume 2, pages 266-273

ISBN: 978-989-758-321-6

presented and discussed in the Section 3; Experimen-

tal results are shown in Section 4; and, ﬁnally, Section

5 presents conclusions and future works.

2 RELATED WORKS

Although there is a considerable number of work re-

lated to facial recognition, few of them present a so-

lution that explores 3D facial characteristics and mor-

phology, in the way that this work does, as will be

explained. Face recognition systems are among the

most reliable biometric systems. They are totally un-

obtrusive and a natural mode of identiﬁcation among

humans (Jain et al., 2004). In well-behaved environ-

ments, the performance can be compared to ﬁnger-

prints (Zaghetto et al., 2017). However, face recogni-

tion is still a challenge, since its accuracy is reduced

due to a number of factors, such as illumination, pose,

distance and many others (Burrows and Cohn, 2009).

Towards improvement of facial recognition sys-

tems, a variety of recent solutions have been pro-

posed. The work of (Haghighat et al., 2016), for

instance, present a fully automatic face recognition

system robust to most common face variations in un-

constrained environments. The system is capable of

recognizing faces from non-frontal views and under

different illumination conditions using only a single

gallery sample for each subject.

Another approach is proposed by (Borgi et al.,

2015), which addresses the problem using a multi-

scale directional framework called Shearlet Network

(SN) to extract facial features, and, a reﬁnement of

the Multi-Task Sparse Learning (MTSL) framework.

One should note that, although previously mentioned

works address the problem of facial recognition, they

do not explore the morphological three-dimensional

(3D) characteristics of faces.

As it is known, 2D images are very sensitive to

illumination changes (Papatheodorou and Rueckert,

2007). Given this fact and the fact that controlling

light in real scenarios are not an easy task, 2D face

recognition biometric systems will never be free from

this weakness. Training algorithms using different il-

lumination scenarios as well as illumination normal-

ization of 2D images has been used, but with limited

success (Papatheodorou and Rueckert, 2007).

In 3D images, however, variations in illumination

do only affect the texture of the image, preserving

its shape and morphological three-dimensional (3D)

characteristics intact (Hesher et al., 2003). To over-

come the problem of illumination on 2D images, the

work of (Chen et al., 2017) proposes that 3D features

may be extracted from 2D images. Based on making

full use of advantages of Sparse Preserving Projec-

tion (SPP) on feature extraction, the discriminant in-

formation was introduced into SPP to arrive at a novel

supervised feather extraction method, that named Un-

correlated Discriminant SPP (UDSPP) algorithm.

Although (Chen et al., 2017) uses 3D features, it

does not work with real 3D images arguing that “al-

though the 3D model method can achieve satisfactory

recognition rate, it needs to pay higher computation

cost”. The work of (Hu et al., 2014) proposed like-

wise a facial recognition method which only uses 3D

images in the face detection process.

Another recent work that addresses the problem

of 3D facial recognition was proposed by (Kim et al.,

2017). In this work, a novel 3D face recognition al-

gorithm using a deep convolutional neural network

(DCNN) and a 3D augmentation technique is pro-

posed. It is mentioned that “training discriminative

deep features for 3D face recognition is very difﬁcult

due to the lack of large-scale 3D face datasets”, so a

CNN is trained on 2D face dataset and then applied to

3D face recognition. Here it should be remarked that

3D facial recognition is still an open ﬁeld to improve-

ment, either because it demands high computational

power or because it lacks of a large dataset to train al-

gorithms or validate results. The work of (Zhou et al.,

2015) proposes a real time 3D face recognition utiliz-

ing a trained two-level cascade classiﬁer and prepro-

cessing the RGB and depth data.

A recent work of (Goswami et al., 2014) proposes

the uniﬁcation of 2D and 3D information in order to

accomplish a hybrid face recognition, applying tech-

niques of entropy and saliency to construct a descrip-

tor and utilizing geometrical analysis of 3D ﬁducial

points. A complete survey of 3D facial recognition is

presented in (Patil et al., 2015).

Table 1 summarizes the comparison between the

related work and the proposed method in this paper.

3 PROPOSED METHOD

In this section, to describe our proposed methodology

all stages are presented in the ﬂuxogram of the Figure

1. The methodology is summarized into four main

steps: 3D Cloud points preprocessing, Face Curva-

ture Maps, Features Extraction and Similarity Match-

ing, all of them described below in Subsections 3.1 up

to 3.4, respectively.

3.1 3D Cloud Points Preprocessing

Firstly, all available 3D data captured contains three-

dimensional information of a full human upper-body.

3D Face Recognition on Point Cloud Data - An Approaching based on Curvature Map Projection using Low Resolution Devices

267

Table 1: Comparison between methods using the characteristics: 3D recognition, real 3D dataset, low Cost sensor, no training,

Low computational power, Point Cloud e Face Curvature Maps (FCM).

Method Characteristics

3D Recog. Real

dataset

Low

cost

sensor

train-

ing

Low

comp.

power

Point

Cloud

FCM

Haghighat, Abdel-Mottaleb,

and Alhalabi (2016)

X X

Borgi, Labate, Elarbi, and

Amar (2015)

X X

Hu N. et al. (2014) X X

Chen Zhanwei, Huang Wei,

and Zhihan Lv (2017)

X X

Kim D. et al. (2017) X X

Guswami, Vatsa and Singh

(2014)

X X X

Zhou, Chen and Wang (2015) X X X X

Proposed method X X X X X X X

Figure 1: Fluxogram of the Proposed Methodology.

In this case, the ﬁrst preprocessing is to deﬁne a

bounded area which includes the face using the Viola-

Jones’ algorithm (Viola and Jones, 2004). Once this

technique was developed for a two dimensional data,

we developed a simple adapted technique that allows

extract the three dimensional data using the two di-

mensional information (vertical and horizontal posi-

tions) from the face, due to a spacial correlation be-

tween 2D (RGB) and 3D (Depth) data. In Figure 1

these steps are described by light-blue color. Samples

of this preprocessing step are described in Figure 2.

Figure 2: Samples results of the 3D preprocessing step -

Subset S

f ace

ICINCO 2018 - 15th International Conference on Informatics in Control, Automation and Robotics

268

All results from this preprocessing step are de-

ﬁned as a subset(S

f ace

) of three-dimensional data

points delimited by the horizontal and vertical spatial

dimensions containing a face of the individual/subject

to be recognized/veriﬁed and described by Equation

f ace

⊂ Ω ∈ R

(1)

where Ω is a set of three dimensional data in spa-

cial domain R

. The subset S

f ace

consists of points

(x,y,z) where k = 1.. . w ... m.

3.2 Face Curvature Maps

Deﬁning M

cov

as the Covariance Matrix of points

(x,y,z), described in Equation 2, which evaluates

covariance around of w three dimensional neighbor

points, as follows:

cov

∑

i=1

− ¯p) · (p

− ¯p)

, (2)

where ¯p is the centroid position of region bounded by

a deﬁned radius r with w

neighbors of set points. The

relationship among the set of p

points and the radius

r is described on Figure 3.

Figure 3: Relationship among the set of p

points and the

radius r.

All face normal curvature indexes (C

) are evalu-

ated by Equations 3 and 4.

cov

· ~v

= σ

· ~v

, j ∈ {0, 1, 2}, (3)

+ σ

(4)

where σ

and ~v

are the eigenvalues and eigenvec-

tor of matrix M

cov

respectively for each p

and σ

σ1 < σ2. In the next step, all extracted C

indexes

are normalized to values between 0(minimal value)

up to 1(maximum value) and described as a face in-

tensity color map, as described in Figure 4-(a) and

(b) as follows. Note that using this projection pro-

cess, all three-dimensional information are included

in a two-dimensional image, allowing to use any clas-

sic face recognition technique, for example the Eigen-

Faces(Belhumeur et al., 1997).

(a)

(b)

(c)

Figure 4: In (a) the Extracted Normal indexes and (b) cur-

vature color maps. For (c) is the normalized to 2D curvature

color map information.

3.3 Features Extraction

From curvature indexes C

all the geometric informa-

tion are carried out of each set of points from a subset

of 3D points clouds. The used features in the match-

ing stage are deﬁned in according to each evaluated

histogram H in Equation 5.

m =

∑

i=1

(5)

3D Face Recognition on Point Cloud Data - An Approaching based on Curvature Map Projection using Low Resolution Devices

269

where b is the previously deﬁned number of bins and

m, as previously deﬁned, is the number of p

points

in a face subset S

f ace

. The variable b deﬁnes how

many features sets will be formed and for high values

of variable b, it implies in the addition of curvatures

information. Otherwise, for reduced values of b, less

curvature details are used in the feature vectors.

3.4 Similarity Matching

In according to (Jain et al., 2004), a biometric sys-

tem is essentially a pattern recognition system that

operates by obtaining biometric data of an individ-

ual, extracting a set of features of the acquired data

and comparing this set of features with the ones al-

ready stored in a database. Depending on the context,

a biometric system can work in a veriﬁcation or iden-

tiﬁcation mode. In the veriﬁcation mode, the system

must validate the identity of an individual comparing

the captured biometric data with the previously cap-

tured and stored data in the database. In the identiﬁ-

cation mode, the system must recognize an individual

by comparing the biometric data with all others pre-

viously stored in the database, searching for the most

similar one.

In this case we are focused on to solve a face veri-

ﬁcation problem using a Similarity Matching schema

presented by (Jain et al., 2004), that can be formally

deﬁned as: Given an input vector of curvature indexes

features C

extracted from the 3D face data and an

alleged identity I, determine if (I,C

) belongs to the

class f

or f

, where f

indicates that the alleged iden-

tity is true and f

that it’s false. C

is compared with

, as the vector of biometric features of the individ-

ual I, to determine its class. Thus

(I,C

) ∈

(

, if S(C

) ≥ t

, otherwise

(6)

where S is a function that measures the similar-

ity score between the vectors C

and C

, and t is the

predeﬁned threshold. S(C

) is called similarity

matching score between the biometric features of the

individual and the alleged identity. The identiﬁcation

problem can be formally deﬁned as: given as entry

a vector of features C

, determine if the identity I

where k ∈ {1, 2, ..., N, N + 1}. Here I

,..., I

are

the identities already in the system and I

N+1

indicates

the rejected case, where no identity is compatible with

the users. Thus

∈







, if max

{S(C

)} ≥ t,k = 1, 2, ..., N

N+1

, otherwise

(7)

where C

is the vector of biometric features cor-

responding to the identity I

, and t is a predeﬁned

threshold.

4 RESULTS

In this Section are presented the evaluation of the pro-

posed methodology, describing details about the used

database and showing results in two speciﬁc tests sce-

narios as follows.

4.1 Database

As previously presented by the ﬂuxogram on Figure

1, the VAP RGB-D Face Database by (Hg et al., 2012)

was used in order to perform the tests scenarios to

evaluate the proposed methodology. This database

provides 31 different subjects, each of them contain-

ing 17 sets of RGB-D images on different poses and

facial expressions (13 poses and 4 facial expressions).

Each pose or expression presents 3 images samples

composed by a RGB color space and a Depth data,

both registered (allowing the correlation of features

between them easily by a translation transform). The

sensor used for data acquisition of this database was

the Microso f t Kinect ﬁrst generation (Zhang, 2012).

All acquired depth images were ﬁltered in order to

treat occlusions and spikes to obtain a smoother and

hole-free Point Cloud data representation of each sub-

ject as described in (Hg et al., 2012).

The main objective of this proposed methodology

is to develop a face recognition algorithm in tests sce-

narios using only faces with the absence of rotation

and occlusion (generated by the face position in rela-

tion to the camera). Speciﬁcally in this case, only the

subject’s frontal face pose was used to allow evaluate

the inﬂuence of all parameters used to adjust the al-

gorithm’s behavior on the facial recognition process.

4.2 Evaluation Process

The evaluation process is composed by two tests sce-

narios, as described as follows:

4.2.1 First Scenario - Biometric Evaluation

In this scenario each subject - in a recognition (classi-

ﬁcation application) system - can be treated as a class.

The used method to provide a possible identiﬁcation

(intra-personal score minimization) is analogue to a

classiﬁer described in Section 3.4. The Equal Error

Rate (EER)(Trentin and Gori, 2001) can be deﬁned

as an objective, threshold-independent measure of the

classiﬁer’s performance for statistical pattern recogni-

tion, which is used to evaluate this classiﬁer and com-

monly used to evaluate biometric systems.

To evaluate and minimize the errors of the pro-

posed methodology, the most inﬂuential variables

ICINCO 2018 - 15th International Conference on Informatics in Control, Automation and Robotics

270

were selected (radius, bins), and then EER was ap-

plied to each one of these variables, deﬁning the opti-

mal value for them.

The rejection criterion established to the ERR was

based in the score obtained from the minimization

function. The ﬁrst criteria deﬁned is the maximum

threshold, avoiding the false recognition from un-

known subjects and refusing badly acquired sensor

outputs. The second criteria is deﬁned by a thresh-

old interval limitation between the two best enroll

matches if they do not belong to the same subject

(since each subject has 3 images samples), allowing

the system to presume doubt between two subjects,

and reject the input.

The radius, responsible for the vicinity description

of a point and, consequently, the curvature intensity

was analyzed in step intervals of 5, generating differ-

ent curvature maps and score for each radius value

from minimization function. In Figure 5 it is possible

to visualize the rates values of false acceptance and

false rejection obtained for each radius values, obtain-

ing an interception at the radius of 26.67, resulting in

an ERR of 3,58% for acceptance and rejection.

Figure 5: Equal Error Rate of Vicinity Radius.

Once the radius is the most independent and the

ﬁrst required variable to be deﬁned in the ﬂuxogram

of the proposed methodology, all remaining variables

will use this optimal radius value ﬁxed as reference to

deﬁne their optimal values from the EER method.

The subsequent analyzed variable is the number

of bins of the intensity histograms obtained from the

curvature maps. The ERR was applied with the same

previously criterion, with radius value ﬁxed during

this analysis, changing only the number of bins used

to represent the histogram as described in Figure 6.

Although it is expected to obtain a better classiﬁer

result using a higher number of bins representation

(and consequently intensity distribution), that doesn’t

guarantee a best discriminative value among subjects.

In Figure 6, values from 8 up to 16 bins achieved an

error of 2,5%, but the number of bins are restricted

to be a power of 2 (to guarantee the equal numeri-

Figure 6: Equal Error Rate of the number of bins.

cal distribution of the intensity values [ranged from

0 to 255]) and located in the exact average of 8 and

16, both of these values are optimal for the number

of bins in this application. The use of values greater

than 16 bins provided a higher disparity between the

images of a same subject, possibly due to the informa-

tion loss (caused by mainly by ﬁltering process) and

the occlusions ﬁlling estimation done by (Hg et al.,

2012) preprocessing, which causes a higher rejection

rate and consequently false rejection as well.

4.2.2 Second Scenario - Accuracy Comparison

In order to obtain a real performance evaluation, this

scenario was developed to compare the performance

of the proposed methodology with others state-of-the-

art techniques related to the used data type and the

face recognition task. This evaluation is focused on

the Rank-1 Accuracy (Lathauwer et al., 2000) of the

facial recognition process, presented as the most usual

evaluation method found in the techniques used in

to comparison. Also, the True Positive Rates (TPR)

and False Positive Rates (FPR) were computed for

both variables Radius and Bins. Receiver Operating

Characteristic (ROC) Curves and Area Under Curves

(AUC) values for accuracy estimates are shown in

Figure 7.

The Rank-1 Accuracy is obtained by a ratio of the

relevant samples from the recognition process (true

positives and true negatives) and the total subject’s en-

rollments in the data base. From extensive tests used

to deﬁne parameters of the proposed methodology,

presented in previously test scenario, the best recogni-

tion results achieved an accuracy of 98, 92% and used

to be compared with the best performance of others

state-of-the-art techniques as described in Table 2.

To obtain a fair comparison, the selected tech-

niques were obtained related to the database present-

ing the identiﬁcation task in facial recognition as well,

3D Face Recognition on Point Cloud Data - An Approaching based on Curvature Map Projection using Low Resolution Devices

271

Table 2: The best Rank 1 Accuracy of the face recognition algorithms related to the VAP RGB-D database.

Data Type Method Rank 1 (%)

RGB Images + Depth Map Goswami et al.(Goswami et al., 2014) 80.6

RGB Images + Depth Image Hu et al. (Hu et al., 2014) 90.0

RGB Images + Depth Image Bormann et al. (Bormann et al., 2013) 96.0

RGB Images + Depth Map Zhou et al. (Zhou et al., 2015) 95.9

Depth Map Saleh and Edirisinghe (Saleh and Edirisinghe, 2016) 96.67

RGB Images + Depth Map Chowdhury et al. (Chowdhury et al., 2016) 98.71

Point Cloud Proposed Methodology 98.92

based on different techniques. In Goswami et al.

(Goswami et al., 2013) a method is proposed to ex-

tract an entropy map from the depth map and the

RGB image of a person and a saliency map from

the RGB image, computing a histogram of gradient

(HOG) from these maps and classifying them by a

Random Forest (RDF). Other work from Goswami et

al. (Goswami et al., 2014) has presented improve-

ments adding a geometric attribute computation from

depth map ﬁducial points, creating the called RISE

(entropy and saliency maps) and ADM (geometric at-

tribute relation) descriptors.

In Hu et al. (Hu et al., 2014) was proposed a

face recognition for a user tracking robotics applica-

tion, using the depth map from head detection and

the RGB image for recognition by illumination nor-

malization, head pose correction and face space pro-

jection. Bormann et al. (Bormann et al., 2013) im-

plements a similar algorithm to Hu et al. algorithm,

Fisherfaces(Belhumeur et al., 1997) space parameter-

ization, a Support Vector Machine (SVM) and Near-

est Neighbor techniques for classiﬁcation. Zhou et

al. (Zhou et al., 2015) proposed a three-dimensional

face recognition using 7 feature points and a two-level

Cascade Classiﬁer, formed by a Decision Tree Classi-

ﬁer in the ﬁrst level, and an improved Euclidian Dis-

tance classiﬁer in the second level. Saleh and Ediris-

inghe proposed an Eigenface-based method, training

models with eigenfaces applied to the normal images

and depth images, under different illumination con-

ditions. Chowdhury et al. (Chowdhury et al., 2016)

proposed a method based on machine learning, that

trains a Neural Network to reconstruct the depth map

from a color image, using the color image and the

real depth map as input elements, and classifying the

reconstructed depth map through another multi class

neural network.

Although close in accuracy performance results

(Figure 7), all considered values above the limits of

excellent (i.e. AUC ≈ 0.90), the estimate for Bin val-

ues (AUC = 0.92) is a 0.02 better than for Radius

vicinity values (AUC = 0.90). All curves are well

over the chance value (i.e. AUC = 0.50).

Figure 7: ROC Curves of Radius and Bins.

5 CONCLUSIONS

This work proposes a methodology based on Point

Cloud face data using curvature map projection for

face recognition. The proposed methodology was

evaluate on two scenarios for biometrics and accu-

racy that describe all features and details about the

parameters setup and performance inﬂuence on face

recognition process.

The Table 2 presents all the Rank-1 Accuracy

obtained in each state-of-the-art technique and con-

cludes that the results obtained by the proposed

methodology outperform the best achieved results

of all the mentioned techniques. This information

demonstrates that the proposed methodology is qual-

iﬁed to be applied in three-dimensional face recogni-

tion problems. In the experiments, the accuracy was

also evaluated plotting Receiver Operating Character-

istic (ROC) Curves and achieving excellent result for

both analyzed variables.

All analysis must be carefully accomplished, since

the implemented techniques in each work are differ-

ent and there are speciﬁcs variations in relation to

the used data base, the used data type and their ex-

perimentations. This allows to archive the conclu-

sion that our proposed methodology presents com-

petitive scores compared to all those techniques. In

ICINCO 2018 - 15th International Conference on Informatics in Control, Automation and Robotics

272

the end, the proposed test scenarios relied on the use

of the VAP RGB-D database (Hg et al., 2012) (due

to its public availability), and the full comparison to

others techniques were impaired due to their experi-

mentations in private data bases. Future works seek

to experiment the proposed methodology in private

databases to obtain broader results.

REFERENCES

Belhumeur, P. N., Hespanha, J. a. P., and Kriegman, D. J.

(1997). Eigenfaces vs. ﬁsherfaces: Recognition using

class speciﬁc linear projection. IEEE Trans. Pattern

Anal. Mach. Intell., 19(7):711–720.

Borgi, M. A., Labate, D., El Arbi, M., and Amar, C. B.

(2015). Sparse multi-stage regularized feature learn-

ing for robust face recognition. Expert Systems with

Applications, 42(1):269–279.

Bormann, R., Zw

olfer, T., Fischer, J., Hampp, J., and

agele, M. (2013). Person recognition for service

robotics applications. In Humanoid Robots (Hu-

manoids), 2013 13th IEEE-RAS International Confer-

ence on, pages 260–267. IEEE.

Bowyer, K. W., Chang, K., and Flynn, P. (2006). A survey

of approaches and challenges in 3d and multi-modal

3d+ 2d face recognition. Computer vision and image

understanding, 101(1):1–15.

Burrows, A. M. and Cohn, J. F. (2009). Anatomy of face. In

Encyclopedia of Biometrics, pages 16–23. Springer.

Chen, Z., Huang, W., and Lv, Z. (2017). Towards a face

recognition method based on uncorrelated discrimi-

nant sparse preserving projection. Multimedia Tools

and Applications, 76(17):17669–17683.

Chowdhury, A., Ghosh, S., Singh, R., and Vatsa, M. (2016).

Rgb-d face recognition via learning-based reconstruc-

tion. In Biometrics Theory, Applications and Systems

(BTAS), 2016 IEEE 8th International Conference on,

pages 1–7. IEEE.

Goswami, G., Bharadwaj, S., Vatsa, M., and Singh, R.

(2013). On rgb-d face recognition using kinect.

In Biometrics: Theory, Applications and Systems

(BTAS), 2013 IEEE Sixth International Conference

on, pages 1–6. IEEE.

Goswami, G., Vatsa, M., and Singh, R. (2014). Rgb-d face

recognition with texture and attribute features. IEEE

Transactions on Information Forensics and Security,

9(10):1629–1640.

Haghighat, M., Abdel-Mottaleb, M., and Alhalabi, W.

(2016). Fully automatic face normalization and sin-

gle sample face recognition in unconstrained environ-

ments. Expert Systems with Applications, 47:23–34.

Haykin, S. and Network, N. (2004). A comprehensive foun-

dation. Neural Networks, 2(2004):41.

Hesher, C., Srivastava, A., and Erlebacher, G. (2003).

A novel technique for face recognition using range

imaging. In Signal processing and its applications,

2003. Proceedings. Seventh international symposium

on, volume 2, pages 201–204. IEEE.

Hg, R., Jasek, P., Roﬁdal, C., Nasrollahi, K., Moeslund,

T. B., and Tranchet, G. (2012). An rgb-d database us-

ing microsoft’s kinect for windows for face detection.

In Signal Image Technology and Internet Based Sys-

tems (SITIS), 2012 Eighth International Conference

on, pages 42–46. IEEE.

Hu, N., Bormann, R., Zw

olfer, T., and Kr

ose, B. (2014).

Multi-user identiﬁcation and efﬁcient user approach-

ing by fusing robot and ambient sensors. In Robotics

and Automation (ICRA), 2014 IEEE International

Conference on, pages 5299–5306. IEEE.

Jain, A., Hong, L., and Pankanti, S. (2000). Biometric iden-

tiﬁcation. Communications of the ACM, 43(2):90–98.

Jain, A. K., Ross, A., and Prabhakar, S. (2004). An intro-

duction to biometric recognition. IEEE Transactions

on circuits and systems for video technology, 14(1):4–

20.

Kim, D., Hernandez, M., Choi, J., and Medioni, G.

(2017). Deep 3d face identiﬁcation. arXiv preprint

arXiv:1703.10714.

Lathauwer, L. D., Moor, B. D., and Vandewalle, J. (2000).

On the best rank-1 and rank-(r1,r2,. . .,rn) approxima-

tion of higher-order tensors. SIAM J. Matrix Anal.

Appl., 21(4):1324–1342.

Papatheodorou, T. and Rueckert, D. (2007). 3d face recog-

nition. In Face Recognition. InTech.

Patil, H., Kothari, A., and Bhurchandi, K. (2015). 3-d face

recognition: features, databases, algorithms and chal-

lenges. Artiﬁcial Intelligence Review, 44(3):393–441.

Saleh, Y. and Edirisinghe, E. (2016). Novel approach to

enhance face recognition using depth maps. In Sys-

tems, Signals and Image Processing (IWSSIP), 2016

International Conference on, pages 1–4. IEEE.

Soldera, J., Schu, G., Schardosim, L. R., and Beltrao, E. T.

(2017). Facial biometrics and applications. IEEE In-

strumentation & Measurement Magazine, 20(2):4–10.

Trentin, E. and Gori, M. (2001). A survey of hybrid

ann/hmm models for automatic speech recognition.

Neurocomputing, 37(1):91–126.

Viola, P. and Jones, M. J. (2004). Robust real-time face

detection. Int. J. Comput. Vision, 57(2):137–154.

Zaghetto, C., Aguiar, L. H. M., Zaghetto, A., Ralha, C. G.,

and de Barros Vidal, F. (2017). Agent-based frame-

work to individual tracking in unconstrained environ-

ments. Expert Systems with Applications, 87:118–

128.

Zhang, Z. (2012). Microsoft kinect sensor and its effect.

IEEE MultiMedia, 19(2):4–10.

Zhou, W., Chen, J.-x., and Wang, L. (2015). A rgb-d face

recognition approach without confronting the cam-

era. In Computer and Communications (ICCC), 2015

IEEE International Conference on, pages 109–114.

IEEE.

3D Face Recognition on Point Cloud Data - An Approaching based on Curvature Map Projection using Low Resolution Devices

273