Upper Limb Anthropometric Parameter Estimation through

Convolutional Neural Network Systems and Image Processing

Andres Guatibonza

, Leonardo Solaque

, Alexandra Velasco

and Lina Pe

nuela

Militar Nueva Granada University, Bogot

a, Colombia

Keywords:

Estimation, Convolutional Neural Networks, Anthropometry, Upper Limb, Image Processing.

Abstract:

Anthropometry is a versatile tool for evaluating the human body proportions. This tool allows the orientation

of public health policies and clinical decisions. But in order to optimize the obtaining of anthropometric mea-

surements, different methods have been developed to determine anthropometry automatically using artiﬁcial

intelligence. In this work, we apply a convolutional neural network to estimate the upper limb’s anthropomet-

ric parameters. With this aim, we use the OpenPose estimator system and image processing for segmentation

with U-NET from a complete uncalibrated body image. The parameter estimation system is performed with

total body images from 4 different volunteers. The system accuracy is evaluated through a global average

percentage of 71% from the comparison between measured values and estimated values. A ﬁne-tuning of

algorithm hyper-parameters will be used in future works to improve the estimation.

1 INTRODUCTION

Anthropometry is a technique that allows to analyze

the physical body features of each person and how

affects their performance. It is usually applied in

different areas such as nutrition, sports, clothing, er-

gonomics, and architecture. It is also commonly used

to study diseases and to assess the nutritional status

of the person (Tov

ee, 2012; Eaton-Evans, 2013). The

anthropometric study makes it possible to calculate a

series of measurements such sectional lengths, thick-

nesses, proportions, weights and sizes in order to ob-

tain information about the individual’s physical and

nutritional status, which allows treating, as the case

may be, certain deﬁciencies or physical aptitudes (Ca-

ballero, 2013; Gallagher et al., 2013).

Anthropometry undoubtedly varies between each

person (Tov

ee, 2012), and the measurement effective-

ness of traditional instruments is unquestionable, but

when measuring, it usually takes some time, this al-

lows the optimization of certain processes like this to

be improved and automated. This fact has led to ex-

ploring forms of parametric estimation of anthropom-

etry and how to develop automatic estimation sys-

https://orcid.org/0000-0001-6102-563X

https://orcid.org/0000-0002-2773-1028

https://orcid.org/0000-0001-7786-880X

https://orcid.org/0000-0002-1925-9296

tems, such as computational models based on neural

networks.

In this context, one of the most popular appli-

cations for human parametric estimation is the pose

estimation, and therefore, the estimation of anthro-

pometry (Agha and Alnahhal, 2012; Ayma et al.,

2016; Chang and Wang, 2015; Dama

sevi

cius et al.,

2018). Works related to development of techniques

for pose estimation have been widely tackled in liter-

ature. For example, in (Batchuluun et al., 2018) con-

volutional neural networks are used for human identi-

ﬁcation based on body movements. Moreover, (Brau

and Jiang, 2016) proposes a deep convolutional neural

network for estimating 3D human pose from monocu-

lar images with 2D annotations. In (Cao et al., 2017),

a efﬁcient detection of pose in 2D of several peo-

ple in an image is carried out using non-parametric

representation that allows to associate parts of the

body with individuals in the image. A human body

posture recognition algorithm based on neural net-

works is developed in (Hu et al., 2016), using the

concept of wireless body area networks (WBAN). In

(Li et al., 2014), a simultaneous heterogeneous sys-

tem of a human pose regressor and detectors of artic-

ulation points and sliding window body parts is pro-

posed in a deep network architecture for the estima-

tion of human pose from monocular images. In (Tang

et al., 2019), a new method is proposed for estimat-

Guatibonza, A., Solaque, L., Velasco, A. and Peñuela, L.

Upper Limb Anthropometric Parameter Estimation through Convolutional Neural Network Systems and Image Processing.

DOI: 10.5220/0010566505830591

In Proceedings of the 18th International Conference on Informatics in Control, Automation and Robotics (ICINCO 2021), pages 583-591

ISBN: 978-989-758-522-7

583

Figure 1: Simpliﬁed diagram of anthropometric parameter estimation.

ing the human pose in 3D from color images and in-

depth images through an RGBD camera using convo-

lutional neural networks. In (Zhu et al., 2017), an es-

timation of the human pose is performed in ﬁxed im-

ages, using a multiple resolution convolutional neu-

ral network (MR-CNN) to train and learn the charac-

teristics of multiple scales of each part of the body.

In (Zhang et al., 2019), the study of pose estima-

tion is mainly focused on the dynamics of the human

limbs for the recognition of actions based on skeletons

by means of a graph edge convolutional neural net-

work (CNN). For some speciﬁc applications, (Farulla

et al., 2016) presents the estimation based on artiﬁ-

cial vision for robot-guided hand tele-rehabilitation,

or the work proposed by (Chen et al., 2018), where

a surveillance human posture dataset (SHPD) based

on more specialized human pose reference points is

presented for surveillance tasks, and (Liu et al., 2018)

where a comprehensive pose network (IntePoseNet)

is proposed to assess the level of customer interest in

a business’ merchandise.

These studies lead to the integration of new meth-

ods or techniques associated with the estimation of

anthropometric indices. These indices are usually ob-

tained through body metrics or using anthropometric

measurement ratios to estimate human proportions or

complementary data that can be obtained from the es-

timation of human pose or other methods with neu-

ral networks as in (Agha and Alnahhal, 2012), where

a neural network and multiple linear regressions are

developed to predict the dimensions of school chil-

dren for an ergonomic design of school furniture.

Moreover, in (Ayma et al., 2016) an approach to

carry out nutritional evaluations to children under ﬁve

years of age, through a system focused on the estima-

tion of anthropometric indices, using image process-

ing techniques and neural network is presented. In

(Chang and Wang, 2015) a learning-based algorithm

is proposed for estimating body shape using multi-

view clothing images.

In the same way, other types of parameters can

be obtained; for instance, in the work proposed by

(Dama

sevi

cius et al., 2018) an automatic system is

developed by using 3D anthropometric measurements

of human subjects for gender detection with k-nearest

neighbors (KNN), Support Vector Machines (SVM)

and neural networks classifying parameters. Authors

in (Rativa et al., 2018) present a complete study of the

application of learning models to estimate height and

weight from anthropometric measurements using sup-

port vector regression, Gaussian process and artiﬁcial

neural networks, as well as in (Saraﬁanos et al., 2016)

a regression-based method is proposed to use privi-

leged information (LUPI) to estimate height using hu-

man metrology. Anthropometric-based true height es-

timation using multilayer perceptron ANN architec-

ture in surveillance areas is tackled in (Sriharsha and

Alphonse, 2019). All these proposed techniques gen-

erate new alternatives to improve or propose estima-

tion strategies with several applications.

In this paper, a study of the application of con-

volutional neural networks and image processing to

estimate anthropometric proportions is presented. We

have developed a method based on neural networks

and pre-calculated systems, such as OpenPose and U-

NET, that allows optimizing anthropometric measure-

ment processes in time and without using measure-

ment instruments. The process is carried out from

image analysis. This study aims to estimate anthropo-

metric parameters from a single uncalibrated image,

using convolutional neural networks for pose estima-

tion and image processing. In this way, we obtain the

lengths and widths of the upper limb. We apply neu-

ral networks and the Euclidean minimization Random

Search technique (EMRS) to estimate the real anthro-

pometric parameters of a person’s upper limb. This

work is limited to the anthropometric estimation of

the upper limb, but it can be applied to the entire hu-

man body. This kind of neural networks models cre-

ates new opportunities for applications related to hu-

man anthropometry in safety and health.

This document is organized as follows: section II

presents the methodology followed to obtain the esti-

mation of human pose, anthropometric lengths, thick-

nesses and perimeters, and a description of the com-

plete system that includes the networks used, and the

ICINCO 2021 - 18th International Conference on Informatics in Control, Automation and Robotics

584

techniques of measurement for the estimation. In sec-

tion III we present the experimental results obtained

by the application of this algorithm to the images of

4 different healthy volunteers. We compare the re-

sults with the actual anthropometric measurements of

the participant. Finally we present, in Section IV, the

conclusions of the work carried out and some future

actions.

2 METHODOLOGY

A neural network system is developed for the acqui-

sition of people’s anthropometric data. The system

estimates anthropometric parameters from an uncali-

brated image. The ﬁrst step is resizing and calibrating

the image. For this, an algorithm based on convolu-

tional neural networks for pose estimation ﬁnds the

points of the joints. Then, the limb lengths are cal-

culated through the Euclidean distance between the

points that make up the body structure. Finally, the

original image is processed to ﬁnd the contours and

segmentation of the person to isolate it from the envi-

ronment. Additionally, other measurements such as

perimeters and thicknesses are calculated using the

EMRS technique. In summary, the system is com-

posed of 3 main phases:

1. Estimation of the human pose using convolutional

networks from an input image.

2. Estimation of the anthropometric lengths from the

keypoints obtained in the convolutional network.

3. Estimation of thickness of the upper limb joints

using image processing to obtain segmental con-

tours of the person.

2.1 Application of Convolutional

Networks for Human Pose

Estimation

The pose estimation is formulated as a CNN-based

regression problem at the body joints. It consists on

detecting the coordinates of a 2D pose (x, y) for each

joint from a RGB image. This pose estimation has

been heavily reshaped by CNN since the introduc-

tion of ”DeepPose” by (Toshev and Szegedy, 2014).

As part of the development of this work, the method

proposed by OpenPose for the estimation of human

posture is used with a non-parametric representation

called Part Afﬁnity Fields (PAF), to ”connect” each

of the body joints found in an image. In essence, the

model consists of pieces based on heat maps for an

approximate location, and a module to sample and

Figure 2: Concept of random search of distance minimiza-

tion.

clip the convolution characteristics at a speciﬁed point

(x,y) for each joint as shown in Fig. 1, as well as

an additional convolutional model for ﬁne adjustment.

Speciﬁcally, this system is a pose machine (Wei et al.,

2016a) that takes as input a color image, and produces

the 2D locations of anatomical keypoints. This sys-

tem consists of a feed-forward network that predicts a

set of 2D conﬁdence maps of body part locations and

a set of 2D vector ﬁelds of part afﬁnity ﬁelds (PAFs),

which encode the degree of association between parts.

The architecture iteratively predicts the afﬁnity ﬁelds

that encode the association between parts, and the de-

tection conﬁdence maps. The iterative prediction ar-

chitecture reﬁnes the predictions in successive stages

with intermediate supervision at each stage. If the

reader wants more information about the operation of

Openpose, please refer to (Cao et al., 2019; Wei et al.,

2016a; Simon et al., 2017; Cao et al., 2017; Wei et al.,

2016b)

In our case, the estimation of the human pos-

ture was limited to the detection of a single person

in the image due to the focus of this work. We

use the COCO model from Microsoft (Lin et al.,

2014), which includes an extensive default dataset of

118.000 images for training and 5.000 images for val-

idation that can be used for object detection, segmen-

tation, and captioning dataset, among other features.

For the COCO model, 14 points are produced, as-

sociated as follows: p

= forehead, p

= Chest, p

Right shoulder, p

= Right elbow, p

= right wrist,

= Left shoulder, p

= Left elbow, p

= Left wrist,

= Right thigh, p

= Right knee, p

= Right an-

kle, p

= Left thigh , p

= Left knee, p

= Left

ankle. Once the image is passed to the model, the net-

work makes a prediction of the coordinates where the

previously deﬁned points are located. The model pro-

duces conﬁdence maps and afﬁnity maps of the pieces

that would be concatenated. Then it veriﬁes whether

Upper Limb Anthropometric Parameter Estimation through Convolutional Neural Network Systems and Image Processing

585

each keypoint belongs to the image. The keypoint lo-

cation is obtained by ﬁnding its maximum value on

the conﬁdence map, and also, includes an algorithm

to ﬁnd false detections, for more information please

refer to (Cao et al., 2019).

2.2 Estimation of Anthropometric

Lengths and Thicknesses of the

Upper Limb from Estimation of

Human Pose and Image Processing

(Regionprops)

Once the coordinates of the body’s joints are obtained

with the pose estimation, the extraction of the charac-

teristics related to the upper limb is carried out. We

consider the distance from shoulder to elbow SEd,

elbow to wrist EW d, and the biacromial width Bw.

These characteristics can be obtained by calculating

the Euclidean distance between two points in pixels

expressed by

x(i+1)

− p

x(i)

)

+ (p

y(i+1)

− p

y(i)

)

, (1)

Where d

is the Euclidean distance between two

points for i = 1, .., 8 upper limb joints; (P

i, P, yi) are

the cartesian position coordinates of the current key-

point, and (P

(i + 1), P, y(i + 1)) are the cartesian po-

sition coordinates of the next adjacent keypoints.

Regarding the characteristics associated with the

thicknesses and the perimeters, such as the wrist

width W w, the arm perimeter Ap, the forearm perime-

ter F p and the elbow width Ew, we obtain them

through image processing. A common way to per-

form segmentation is through a Fully Convolutional

Network (FCN), better known as U-NET. In the same

way, this network has pre-trained models with an ex-

tensive image dataset of people with pointers or iden-

tiﬁers to perform the segmentation. The result ob-

tained at the output of the network is a mask of the

person excluding the background. This approach to

image processing will allow to segment it into re-

gions. Similarly, if the reader wants more information

about the operation of U-NET, please refer to (Ron-

neberger et al., 2015).

To obtain the widths and perimeters, we start from

the original image M

original

which is entered into pre-

trained U-NET model where the matrix of the identi-

ﬁed person mask is obtained at the output M

mask

. Ad-

ditionally, noise ﬁltering is performed to soften the

edges of the mask and eliminate unwanted disper-

sion. Filtering is done with a morphological transfor-

mation since the image can be described as a binary.

The morphological operators used for the decision are

erosion and threshold. A kernel array H(u, v) slides

through the image (as in 2D convolution) where 1 is

assigned if all pixels below the kernel are greater than

the threshold; otherwise it erodes (reduces to zero).

The matrix resulting from ﬁltering M

f ilter

can be ob-

tained by

f ilter(i, j)

= M

mask(i, j)

∗ H(u, v)

∑

i=0

∑

j=0

f ilter(i, j)

(

1 if M

f ilter(i, j)

> threshold

0 if M

f ilter(i, j)

≤ threshold

(2)

where M

f ilter(i, j)

is the resulting matrix after the ﬁl-

tering; H(u, v) is the kernel matrix that performs the

convolution across the image and evaluates with re-

spect to the threshold. The selection of the threshold

is free, typically, a value 0.5 is used.

Then, the background is eliminated by subtracting

the matrix belonging to the original image M

original

from the matrix resulting from the ﬁltered mask

f ilter

. After that, a blur ﬁlter is applied to the binary

image to eliminate the residual noise and preserve the

smoothness of the edges. To do this, we apply the im-

age convolution. Then, we take the average of all the

pixels below the kernel area, to replace the central ele-

ment with this average. We use cv2.blur () function of

OpenCV with this purpose. For more information of

this library the reader is encouraged to review (Mord-

vintsev and K, ).

From the binarized image, the contours of the per-

son can be easily obtained by evaluating the matrix

of pixels when a change from 0 to 1 is detected, and

viceversa. Each individual contour is a matrix of co-

ordinates (x,y) of the object border points. These con-

tours are the limits of a shape with the same intensity.

We use OpenCV function cv2.ﬁndContours () with

this purpose. This function generates contours elimi-

nating redundant points and keeping those that offer a

great amount of information to save processing mem-

ory. The stored points that belong to the contour are

included in the original image.

Subsequently, the Canny edge detection algorithm

(Canny, 1986) is applied. This algorithm ﬁrst removes

noise in the image with a 5x5 Gaussian ﬁlter, since

edge detection is susceptible to noise in the image.

Then, the intensity of the gradient is found for hor-

izontal, vertical and diagonal detection at the edges

of the ﬁltered image. Sobel’s edge detection operator

(Patnaik and Yang, 2012) for example, returns a value

for the ﬁrst derivative in the horizontal direction (Gx)

and the vertical direction (Gy). Then, the edge gradi-

ICINCO 2021 - 18th International Conference on Informatics in Control, Automation and Robotics

586

ent and direction are determined by

EdgeGradient(G) =

+ G

Angle(θ) = tan

−1





(3)

The pixels of the edges are kept or eliminated by

the hysteresis threshold in the magnitude of the gradi-

ent. In our case, we used a Gaussian width σ = 1. The

function skimage.feature.canny () of Scikit-image li-

brary performs this detection (Canny, 2020).

After calculating the edges, an exact Euclidean

distance transform is used to evaluate the distances

between the points obtained from the edges and their

correlation using a heat map. For this, we used

the function distance transform edt() of SciPy (Scipy,

2020).

Using the resulting distance map, we calculate

the coordinates of the local maxima found in the im-

age. If the peaks are ﬂat, that is, several adjacent

pixels have identical intensities, the coordinates of

all those pixels are returned as local maxima. The

calculation of local maxima is be deﬁned by a re-

gion of 2 ∗ mindistance + 1 where the selection of

mindistance will deﬁne the amount of points that

the algorithm manages to ﬁnd. We use the func-

tion feature.peak local max() of scikit-image library

(OpenCV, 2020b).

Then, with the deﬁnition of the local maxima,

edge segmentation is used where the limits between

the highlighted labeled regions are applied to the

original image. We use the function segmenta-

tion.mark boundaries() of scikit-image library for this

purpose (OpenCV, 2020a). Subsequently, the mark-

ers obtained by the segmentation will serve to ﬁll in

the areas enclosed by the segmentation, evaluating the

color values of the pixels in each area including the

background, and then calculating the average values

to normalize the segmented area. The background of

the image is normalized by assigning a user-deﬁned

color that represents a differentiation with the person

to avoid confusing the colors between the person and

the background.

Then, the thickness and the perimeters are esti-

mated on the resulting image using a random search

technique to minimize the euclidean distances. This

technique consists in selecting the joint and mid-

point between the joints and evaluating the euclidean

distance between a series of random points located

outside the person’s space and the chosen reference

point. The workspace radius that restricts this evalua-

tion is deﬁned as shown in Fig. 2.

The orientation of the points near A and B to the

reference point R determine the arms orientation in

the photo (raised or relaxed) and whether it is the right

arm or the left arm. Then we deﬁne the radius of the

workspace on which the search is carried out by using

r = 0.5 ∗

− B

) + (A

− B

)

Subsequently, a series of random points is gener-

ated on the radial space, and each distance towards the

reference point is evaluated iteratively until obtaining

the smallest possible distance that occurs when the

edge of the person is reached. In this way, we ob-

tain the estimates of the thicknesses and the perime-

ters of the upper limb. Then it is necessary to make a

conversion of the estimates calculated from pixels to

centimeters as

ExtractedLength

= α ∗ ExtractedLength

(4)

where α is a proportionality constant associated with

a known horizontal distance between the camera and

the person.

3 EXPERIMENTAL RESULTS

For the validation of the estimation system, a set of

images with known anthropometric values are used.

Four volunteers (2 men and 2 women) participated

in the experiment. Their anthropomorphic measure-

ments were previously obtained with measurement in-

struments. We used rulers to measure distances and

calipers to measure thicknesses. Due to current san-

itary restrictions, we worked with a limited number

of volunteers, however, in future work we will extend

the participation to give a greater validity to the esti-

mation method.

The images are inserted into the subsystem to

obtain the pose estimation, calculating the coordi-

nates (x,y) associated with the joints and other de-

ﬁned points. The images were taken in uncontrolled

environments to validate the effectiveness of the al-

gorithm. For this reason, in this paper we present the

cases of four volunteers in a non-controlled environ-

ment, with raised arms, which is a particular exam-

ple where the person is doing any pose gesture and

the background is not one color. In Fig.3 we show

the resulting estimation of the deﬁned points and the

association of these points to each joint. The input

image has a resolution of 500x500 pixels. The algo-

rithm returns the parts that belong to the background

and those that belong to the person’s body to place the

points of the joints. Some points of the lower limbs

are not fully centered in the human body joints due to

the similarity of colors with the background. Then,

the lengths in pixels of SEd, EW d and Bw are calcu-

lated using (1). The proportionality of distances cal-

culated in the image in Fig. 3 keeps a certain consis-

tency if it is compared with the proportionality of the

Upper Limb Anthropometric Parameter Estimation through Convolutional Neural Network Systems and Image Processing

587

Figure 3: Sample image pose estimation.

Figure 4: Segmentation process and image processing.

Table 1: Measured vs estimated anthropometric values.

Volunteer 1 Volunteer 2 Volunteer 3 Volunteer 4

Parameter

Measured

(cm)

Estimated

(cm)

Measured

(cm)

Estimated

(cm)

Measured

(cm)

Estimated

(cm)

Measured

(cm)

Estimated

(cm)

SEd 35 36,6 33 38,4 30 35 24 31,3

EW d 26 31,5 27 38 26 35,6 24 27,1

Bw 42 41,4 40 53,4 38 47,8 38 42,8

W w 5,7 4,32 5,6 5,1 5,4 8,6 5,7 3,7

Ap 30 33,06 35 73 33 34,9 35 21,6

F p 25 21,18 25 10,5 20 10,7 28 32,1

Ew 8,9 5,64 8,7 10 8,9 7 8,9 3,7

ICINCO 2021 - 18th International Conference on Informatics in Control, Automation and Robotics

588

Figure 5: Shortest distance search in processed image.

anthropometric data in (Avila-Chaurand et al., 2007).

Proportionality must maintain a ﬁxed value for all the

measured parameters. Table. 2 shows that SEd and

EW d have a ﬁxed proportionality value of 0.3± 0.025

but Bw differs in 0.15 units approximately, this is due

to the fact that the estimation of the points that belong

to the shoulders’ biacromial width are not accurately

located.

Table 2: Proportionality estimation values.

Parameter Estimated

(px)

Measured

(cm)

Proportion

SEd 56 35 0.375

EW d 40 26 0.35

Bw 35 42 0.2

Then, we calculate the thicknesses and the

perimeters using the original image for the segmen-

tation system with U-NET. The result of the esti-

mation is shown in Fig.4.b where the identiﬁed ma-

trix of the person mask M

f ilter

, with noise ﬁltering

and edge smoothing using a threshold of 0.5 is ob-

tained. By subtracting the matrix M

original

from the

f ilter

we eliminate the background, as shown in

Fig.4.c. In addition to matrix M

f ilter

, a blur ﬁlter is

applied to eliminate residual noises that may have re-

mained in the ﬁrst ﬁltering. The idea is to approx-

imate the contour of the person as closely as possi-

ble so the estimation can be improved. The result

of the second ﬁltering is shown in Fig.4.d. Subse-

quently, the outlines of the binary and ﬁltered im-

age are found using regionprops and combined with

the original image as shown in Fig.4.e. Then, the

characteristic edges and strokes found throughout the

body are calculated with Canny’s edge detection al-

gorithm (Canny, 1986) using a Gaussian width σ = 1,

as shown in Fig.4.f. The correlation of the edges and

distances between the points obtained from the edge

are shown in Fig.4.g, where later in Fig.4.h-i, the lo-

cal max peak distances and IDs are calculated using

a region with mindistance = 5 . Then, the segmen-

tation of the points obtained is carried out and the

segmented areas are repainted with the average RGB

value as shown in Fig.4.j-k. The background color

has on average a bluish color so when repainting the

areas, the background is also repainted. Finally, the

background is normalized by assigning a RGB value

of (255,255,255) corresponding to white color, thus

resulting in the image of Fig.4.l as the output image

for estimating thickness and perimeters.

Now, we calculate the thicknesses and the perime-

ters W w, Ap, F p and Ew. For this, we identify the ori-

entation of the points A and B with the slope of those

points and we evaluate the highest and lowest values

of the coordinates in (x, y) in the cartesian plane of

the image as shown in Fig.5.a. Then, we deﬁne the

searching radius (red circle), and we evaluate the mi-

nor euclidean distance performing the random search

algorithm with 100 points as shown in Fig.5.b.

Then, the obtained anthropometric proportions es-

timation results are converted from px to cm with

α = 0.6. This procedure was repeated for the other

volunteers. Table 1 shows a comparison of the esti-

mated and measured data. In average, the accuracy is

71% between the measured values and the estimated

values. SE

is the parameter with the highest aver-

age accuracy (83%), while A

is the parameter with

the lowest average accuracy (59%). Volunteer 1 has

the highest average accuracy (84%), while volunteer

2 has the lowest average accuracy (60%). The accu-

racy can be increased using an image with a more

controlled environment. On the other hand, the res-

olution of the camera, the image processing indices

and the number of points in the EMRS technique can

also improve the estimation.

4 CONCLUSIONS

In this paper, we presented a system for the estima-

tion of anthropometric parameters. This system was

implemented through the use of convolutional neural

networks and image processing. We proposed a sys-

tem based on 3 phases: estimation of the human pose,

estimation of anthropometric lengths and estimation

of thickness using image processing.

Upper Limb Anthropometric Parameter Estimation through Convolutional Neural Network Systems and Image Processing

589

Human pose estimation is made from an image of

total body. Four volunteers participated in this study.

A particular example of an uncalibrated image was

analyzed. This image was taken in a random environ-

ment where the pose estimation algorithm ﬁnds ad-

equately the joint points. Then, the anthropometric

lengths are calculated from the coordinates obtained

from the human pose by means of a euclidean cal-

culation. Then, we obtained the thicknesses and the

perimeters of the upper limb through segmentation,

image processing and a Euclidean minimization ran-

dom search (EMRS). The results obtained were com-

pared with measurements previously taken with rulers

for the distances and calipers for the thicknesses, get-

ting a global average accuracy of 71% between the

measurements in all the subjects. We deﬁne empir-

ically the hyperparameters, but other strategies can

be proposed in the future to make a more accurate

ﬁne-tunning of these hyperparameters in order to get

better results. This method promises to optimize the

estimation of anthropometric measurements automat-

ically without using other instrumentation, only from

an image and the distance between the camera and the

person.

Currently we are performing more tests of the es-

timation system, considering a higher number of par-

ticipants. In this way we will have different anthropo-

metric proportions in order validate and generalized

the algorithm by ﬁne-tuning the hyperparameters of

the estimation. In the future we will also general-

ize the estimation to obtain other data such as muscle

strength to use it for example, in rehabilitation appli-

cations.

ACKNOWLEDGEMENTS

This work is supported by Universidad Militar Nueva

Granada- Vicerrectoria de Investigaciones, under re-

search project IMP-ING-3127, entitled ’Dise

no e im-

plementaci

on de un sistema rob

otico asistencial para

apoyo al diagn

ostico y rehabilitaci

on de tendinopat

ıas

del codo’.

REFERENCES

Agha, S. R. and Alnahhal, M. J. (2012). Neural network

and multiple linear regression to predict school chil-

dren dimensions for ergonomic school furniture de-

sign. Applied Ergonomics, 43(6):979–984.

Avila-Chaurand, R., Prado-Le

on, L., and Gonz

alez-Mu

noz,

E. (2007). Dimensiones antropom

etricas de la

poblaci

on latinoamericana : M

exico, Cuba, Colom-

bia, Chile / R. Avila Chaurand, L.R. Prado Le

on, E.L.

Gonz

alez Mu

noz.

Ayma, V. A., Ayma, V. H., and Torre, L. G. A. (2016). Nu-

tritional assessment of children under ﬁve based on

anthropometric measurements with image processing

techniques. In 2016 IEEE ANDESCON. IEEE.

Batchuluun, G., Naqvi, R. A., Kim, W., and Park, K. R.

(2018). Body-movement-based human identiﬁcation

using convolutional neural network. Expert Systems

with Applications, 101:56–77.

Brau, E. and Jiang, H. (2016). 3d human pose estimation

via deep learning from 2d annotations. In 2016 Fourth

International Conference on 3D Vision (3DV). IEEE.

Caballero, B. (2013). Nutritional assessment: Clinical

examination. In Encyclopedia of Human Nutrition ,

pages 233–235. Elsevier.

Canny, J. (1986). A computational approach to edge de-

tection. IEEE Transactions on Pattern Analysis and

Machine Intelligence, PAMI-8(6):679–698.

Canny, J. (2020). Canny edge detector. https://scikit-

image.org/docs/dev/auto examples/edges/plot canny.

[Online; accessed october-2020].

Cao, Z., Hidalgo Martinez, G., Simon, T., Wei, S., and

Sheikh, Y. A. (2019). Openpose: Realtime multi-

person 2d pose estimation using part afﬁnity ﬁelds.

IEEE Transactions on Pattern Analysis and Machine

Intelligence.

Cao, Z., Simon, T., Wei, S.-E., and Sheikh, Y. (2017). Real-

time multi-person 2d pose estimation using part afﬁn-

ity ﬁelds. In CVPR.

Chang, W.-Y. and Wang, Y.-C. F. (2015). Seeing through the

appearance: Body shape estimation using multi-view

clothing images. In 2015 IEEE International Confer-

ence on Multimedia and Expo (ICME). IEEE.

Chen, Q., Zhang, C., Liu, W., and Wang, D. (2018).

SHPD: Surveillance human pose dataset and perfor-

mance evaluation for coarse-grained pose estimation.

In 2018 25th IEEE International Conference on Image

Processing (ICIP). IEEE.

Dama

sevi

cius, R., Camalan Seda, Sengul Gokhan, Misra

Sanjay, and Maskeli

unas Rytis (2018). Gender detec-

tion using 3d anthropometric measurements by kinect.

Eaton-Evans, J. (2013). Nutritional assessment: Anthro-

pometry. In Encyclopedia of Human Nutrition, pages

227–232. Elsevier.

Farulla, G. A., Pianu, D., Cempini, M., Cortese, M., Russo,

L., Indaco, M., Nerino, R., Chimienti, A., Oddo, C.,

and Vitiello, N. (2016). Vision-based pose estimation

for robot-mediated hand telerehabilitation. Sensors,

16(2):208.

Gallagher, D., Chung, S., and Akram, M. (2013). Body

composition. In Encyclopedia of Human Nutrition,

pages 191–199. Elsevier.

Hu, F., Wang, L., Wang, S., Liu, X., and He, G. (2016). A

human body posture recognition algorithm based on

BP neural network for wireless body area networks.

China Communications, 13(8):198–208.

Li, S., Liu, Z.-Q., and Chan, A. B. (2014). Heteroge-

neous multi-task learning for human pose estimation

ICINCO 2021 - 18th International Conference on Informatics in Control, Automation and Robotics

590

with deep convolutional neural network. International

Journal of Computer Vision, 113(1):19–36.

Lin, T.-Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ra-

manan, D., Doll

ar, P., and Zitnick, C. L. (2014). Mi-

crosoft COCO: Common objects in context. In Com-

puter Vision – ECCV 2014, pages 740–755. Springer

International Publishing.

Liu, J., Gu, Y., and Kamijo, S. (2018). Integral cus-

tomer pose estimation using body orientation and vis-

ibility mask. Multimedia Tools and Applications,

77(19):26107–26134.

Mordvintsev, A. and K, A. Smooth-

ing images. https://opencv-python-

tutroals.readthedocs.io/en/latest/py tutorials/py imgp

roc/py ﬁltering/py ﬁltering.html.

OpenCV (2020a). Module: segmentation. https://scikit-

image.org/docs/dev/api/skimage.segmentation.html.

[Online; accessed october-2020].

OpenCV (2020b). peak local max. https://scikit-

image.org/docs/0.7.0/api/skimage.feature.

peak.html. [Online; accessed october-2020].

Patnaik, S. and Yang, Y.-M., editors (2012). Soft Computing

Techniques in Vision Science. Springer Berlin Heidel-

berg.

Rativa, D., Fernandes, B. J. T., and Roque, A. (2018).

Height and weight estimation from anthropometric

measurements using machine learning regressions.

IEEE Journal of Translational Engineering in Health

and Medicine, 6:1–9.

Ronneberger, O., Fischer, P., and Brox, T. (2015). U-

net: Convolutional networks for biomedical image

segmentation. In Navab, N., Hornegger, J., Wells,

W. M., and Frangi, A. F., editors, Medical Image Com-

puting and Computer-Assisted Intervention – MICCAI

2015, pages 234–241, Cham. Springer International

Publishing.

Saraﬁanos, N., Nikou, C., and Kakadiaris, I. A. (2016).

Predicting privileged information for height estima-

tion. In 2016 23rd International Conference on Pat-

tern Recognition (ICPR). IEEE.

Scipy (2020). distance transform edt.

https://docs.scipy.org/doc/scipy/reference/genera-

ted/scipy.ndimage.distance transform edt.html.

[Online; accessed october-2020].

Simon, T., Joo, H., Matthews, I., and Sheikh, Y. (2017).

Hand keypoint detection in single images using mul-

tiview bootstrapping. In CVPR.

Sriharsha, K. V. and Alphonse, P. (2019). Anthropometric

based real height estimation using multi layer pecep-

tron ANN architecture in surveillance areas. In 2019

10th International Conference on Computing, Com-

munication and Networking Technologies (ICCCNT).

IEEE.

Tang, H., Wang, Q., and Chen, H. (2019). Research on 3d

human pose estimation using RGBD camera. In 2019

IEEE 9th International Conference on Electronics In-

formation and Emergency Communication (ICEIEC).

IEEE.

Toshev, A. and Szegedy, C. (2014). DeepPose: Human pose

estimation via deep neural networks. In 2014 IEEE

Conference on Computer Vision and Pattern Recogni-

tion. IEEE.

Tov

ee, M. (2012). Anthropometry. In Cash, T., editor, En-

cyclopedia of Body Image and Human Appearance,

pages 23 – 29. Academic Press, Oxford.

Wei, S., Ramakrishna, V., Kanade, T., and Sheikh, Y.

(2016a). Convolutional pose machines. CoRR,

abs/1602.00134.

Wei, S.-E., Ramakrishna, V., Kanade, T., and Sheikh, Y.

(2016b). Convolutional pose machines. In CVPR.

Zhang, X., Xu, C., Tian, X., and Tao, D. (2019). Graph edge

convolutional neural networks for skeleton-based ac-

tion recognition. IEEE Transactions on Neural Net-

works and Learning Systems, pages 1–14.

Zhu, A., Jin, J., Wang, T., and Zhu, Q. (2017). Human pose

estimation via multi-resolution convolutional neural

network. In 2017 4th IAPR Asian Conference on Pat-

tern Recognition (ACPR). IEEE.

Upper Limb Anthropometric Parameter Estimation through Convolutional Neural Network Systems and Image Processing

591