Skin Temperature Measurement based on

Human Skeleton Extraction and Infra-red Thermography

An Application of Sensor Fusion Methods in the Field of Physical Training

Julia Richter

, Christian Wiede

, Sascha Kaden

, Martin Weigert

and Gangolf Hirtz

Department of Electrical Engineering and Information Technology,

Chemnitz University of Technology, Reichenhainer Str. 70, 09126 Chemnitz, Germany

Department of Behavioural and Social Sciences, Chemnitz University of Technology,

Reichenhainer Str. 70, 09126 Chemnitz, Germany

Keywords:

Infra-red Thermography, Skin Temperature, Sensor Fusion, Human Skeleton, Physical Training.

Abstract:

Skin temperature measurements play a vital role in the diagnosis of diseases. This topic is also increasingly

investigated for applications in the ﬁeld of physical training. One of the limitations of state-of-the-art methods

is the manual, time-consuming way to measure the temperature. Moreover, extant literature gives only little

insight into the skin temperature behaviour after the training. The aim of this study was to design an automatic

method to measure the skin temperature during and after training sessions for the biceps brachii. For this pur-

pose, we fused thermal images and skeleton data to locate this muscle. We could successfully demonstrate the

working principle and observed a temperature increase even several minutes after the end of the training. This

study therefore contributes to the automation of skin temperature measurements. A transfer of our approach

could be beneﬁcial for other application ﬁelds, such as medical diagnostics, as well.

1 INTRODUCTION

Infra-red thermography plays an increasingly impor-

tant role in a wide range of application ﬁelds. In

medicine, it is a non-invasive diagnostic method to

detect abnormal body temperatures, which are indi-

cators for a variety of diseases, such as breast cancer

and diabetic vascular disorder (Lahiri et al., 2012) or

arthritis (Ring and Ammer, 2012). In sports medicine,

infra-red thermography is applied to measure the de-

gree of regeneration or to detect overuse reactions in

order to avoid injuries (Hildebrandt et al., 2012).

Recently, researchers have shown an increased in-

terest in the investigation of the relationship between

skin temperature and muscle activity during training.

To date, however, only few studies examined the tem-

perature proﬁle after training sessions. Moreover, the

methods presented in previous work determined the

skin temperature of a region in a manual way and are

therefore time-consuming. The purpose of this study

is to facilitate skin temperature determination of a cer-

tain region by introducing an automatic method. We

fused images from a thermal camera as well as skele-

ton data from the Kinect sensor. For this, we have

developed a calibration method that enables this sen-

sor fusion. We demonstrated the working principle

for the biceps brachii and analysed the heat develop-

ment during and after the training with biceps curls.

Therefore, this study makes a major contribution to

research on the relationship between skin temperature

and muscle activity in the ﬁeld of sports science. At

this point, we assumed that the muscle is not covered

by clothes. The study has been organised as follows:

Section 2 examines the extant literature on infra-red

thermography and calibration. Section 3 begins by

presenting the sensor system and the calibration pro-

cedure, including both intrinic and extrinsic calibra-

tion as well as the mapping process. Subsequently,

this section explains the skin temperature measure-

ment. Section 4 is concerned with the evaluation

methodology, while Section 5 presents and discusses

the results. We close the paper with conclusions and

an outlook at future work in Section 6.

2 RELATED WORK

2.1 Infra-red Thermography

More recent attention has focused on the evaluation

of surface temperature changes in training diagnos-

Richter J., Wiede C., Kaden S., Weigert M. and Hirtz G.

Skin Temperature Measurement based on Human Skeleton Extraction and Infra-red Thermography - An Application of Sensor Fusion Methods in the Field of Physical Training.

DOI: 10.5220/0006095100590066

In Proceedings of the 12th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2017), pages 59-66

ISBN: 978-989-758-227-1

tics. Variations in temperature can indicate the grade

of muscle activity, which can be visualised by evaluat-

ing infra-red thermography images. In the following,

existing literature is reviewed with the focus on the

applied analysis methods.

Bartuzi et al. revealed statistically signiﬁcant cor-

relations between the skin temperature and EMG pa-

rameters for the biceps brachii by employing infra-

red thermography (Bartuzi et al., 2012). They ad-

justed a thermal camera in such a way that the cap-

tured thermal images showed only the muscle itself.

The skin temperature was then determined using a

special software that analysed manually deﬁned re-

gions. Formenti et al. examined the skin temperature

development of the region that covers the calf mus-

cle after standing calf raise exercises (Formenti et al.,

2013). They observed an increasing skin tempera-

ture during and also two minutes after the exercise,

which is an indicator that the active muscle emitted

heat. Their temperature determination works semi-

automatically: they manually selected a region of in-

terest (ROI), which was located on the calf and then

automatically determined the ﬁve warmest pixels. In

the next step, they considered a region of ﬁve by ﬁve

pixels around each of these warmest pixels and av-

eraged them. This average was regarded as the ﬁnal

temperature for the previously selected ROI. Ludwig

et al. thereupon veriﬁed a correlation between this

very method and a method that simply averages the

pixels belonging to the selected ROI (Ludwig et al.,

2014).

Other approaches were based on anatomical infor-

mation, which allowed the derivation of muscle lo-

cations in the thermal image. Two examples are the

studies of Bandeira et al. and Neves et al.: accord-

ing to the arrangement of the different muscles in the

thigh, Bandeira et al. manually selected ROIs for tem-

perature determination (Bandeira et al., 2012). In a

similar manner, Neves et al. estimated the centre of

the biceps brachii by exploiting the knowledge about

the muscle anatomy (Neves et al., 2014).

Several methods employ markers that were at-

tached to the skin. In this way, speciﬁc muscle tem-

peratures were determined by manually evaluating the

region between these markers in the recorded ther-

mal images. Fr

ohlich et al., for example, bonded

corks to a human body in order to measure temper-

atures belonging to various regions (Fr

ohlich et al.,

2014), while Neves et al. ﬁxed tapes to the biceps

brachii. Further signiﬁcant analyses and discussions

on thermographic investigations, their techniques and

inﬂuencing factors were presented in the works of

Fern

andez et al. and Ring and Ammer (Fern

andez-

Cuevas et al., 2015), (Ring and Ammer, 2015).

The evidence presented in this section suggest

that there is need to investigate automatic methods

for skin temperature evaluation. To date, automatic

techniques to measure muscle temperature cannot be

found in extant literature. Therefore, we propose such

an automatic method that employs skeleton extraction

algorithms to detect speciﬁc muscles. This enables an

automatic localisation and analysis of speciﬁc muscle

regions. Moreover, we demonstrated that after per-

forming biceps curls, the temperature of the surface

surrounding the biceps brachii increases, which sup-

ports the ﬁndings of Formenti et al.

2.2 Calibration

In order to combine skeletal data with thermal images,

the Kinect version 1.0 skeleton joint coordinates and

the images of the thermal camera had to be fused. For

this procedure, a calibration of both sensors was re-

quired to allow a mapping of the thermal image to the

Kinect RGB image, in which the skeleton joints are

represented in 2-D coordinates.

The combination of a thermal camera and an RGB

camera or a depth sensor, such as the Kinect, was

widely used for different applications. Thomanek

et al. implemented a pedestrian detection algorithm

based on both RGB images and thermal images

(Thomanek et al., 2011). On the one hand, pedestri-

ans could be detected in thermal images even at night,

when the detection on the RGB image failed. On the

other hand, the detection in thermal image failed on

warm days and under strong solar radiation. In these

cases, however, pedestrians could be detected in the

RGB image. The principle of compensating single

sensor weaknesses was extended by Geschwandtner

et al. (Gschwandtner et al., 2011). They utilised a

thermal camera, a laser range scanner and an RGB

camera in the ﬁeld of autonomous driving. Other

applications of thermal imaging in combination with

other sensors are the maintenance of power equip-

ment (Nakagawa et al., 2014), the building sector (Vi-

das et al., 2013) and e-rehabilitation (Richter et al.,

2016). All these applications required a reliable cal-

ibration to fuse several optical systems to one com-

bined world coordinate system. For this purpose,

three aspects had to be considered: an adequate cal-

ibration target, whose pattern is detectable for every

employed sensor, as well as the intrinsic and the ex-

trinsic calibration.

In the visible light spectrum, a planar checker-

board is the commonly used calibration target. By

detecting the inner checkerboard corners, information

about distortions in the RGB images can be estimated.

However, this working scheme is not applicable for

VISAPP 2017 - International Conference on Computer Vision Theory and Applications

Figure 1: Sensor system with thermal camera and Kinect.

thermal images. Even if a checkerboard with appro-

priate material is illuminated by an infra-red heater,

the edges of the checkerboard appear smoothed and

cannot be accurately localised. To overcome this

problem, alternative calibration targets have been de-

veloped. Nakagawa et al. proposed two plastic plates

with inverse circular patterns (Nakagawa et al., 2014),

which could be plugged into each other. For the cal-

ibration of a thermal and an RGB sensor, one plate

was plugged onto the other and served as an isolating

layer when they were heated. The isolating layer was

then removed, so that only the heated circles were vis-

ible in the thermal image. At the same time, the circle

boundaries could be detected in the RGB image as

well. In comparison to that, Vidas et al. used a planar

mask with cut-out squares, which was held in front

of a backdrop with a different temperature (Vidas

et al., 2013). With this method, checkerboard corners

could more accurately detected than in case of the

commonly used checkerboard calibration described

above. Rangel et al. introduced a similar approach,

but instead of squares, they cut out asymmetrical cir-

cles to increase the accuracy (Rangel et al., 2014).

They heated this pattern to enable the circle detection

in the thermal image. In contrast to the calibration of

RGB sensors, the characteristic of this kind of ther-

mal calibration targets is that the edges are blurred in

the thermal image. The advantage of circles in such

cases is a more stable detection: the centre of a cir-

cles can be detected more robustly than edges in such

cases. While all these approaches used passive cali-

bration targets, Ellmauthaler et al. suggested an LED

grid (Ellmauthaler et al., 2013), whereas Gschwandt-

ner et al. proposed a checkerboard with resistors as

active calibration targets (Gschwandtner et al., 2011).

An overview about different calibration targets was

presented by Rangel et al. (Rangel et al., 2014).

The intrinsic calibration in all presented ap-

proaches was always performed according to the

method of Zhang (Zhang, 2000), which assumed a

pinhole camera model. Sections 3.2 and 3.3 present a

detailed explanation of the intrinsic and extrinsic cal-

ibration procedures we applied in our study.

3 METHODS

This section presents the sensor system, the calibra-

tion steps and the procedure to measure the skin tem-

perature.

3.1 Sensor System

The sensor system comprises an RGB-D sensor, i. e.

a Kinect version 1.0, which additionally provides a

human skeleton stream, and a thermal camera, i. e. a

FLIR A35sc. The thermal camera has a thermal reso-

lution of 50 mK. The spatial resolution of the sensor

is 320×256 pixel. The rotation between both sensors

was as small as possible, so that both sensors share

approximately the same re-projection plane. The sen-

sor system is illustrated in Figure 1.

3.2 Intrinsic Calibration

The aim of the camera calibration is the mapping of

the thermal image onto the Kinect RGB image, in

which the skeleton joints are represented in 2-D im-

age coordinates.

For all calibration steps that are connected with

the thermal sensor, we used an aluminium plate with

cut-out circles, which is similar to Rangel et al.

(Rangel et al., 2014). In contrast to Rangel et al., the

circles on our target are symmetrically arranged, see

Figure 2. Similar to Vidas et al. (Vidas et al., 2013),

we used a TFT monitor as a backdrop that emits a

different temperature, as illustrated in Figure 2. By

using this calibration target, we could detect circles

in the thermal image according to the algorithm by

Suzuki and Abe (Suzuki and Abe, 1985).

The thermal camera was intrinsically calibrated

using the method proposed by Zhang (Zhang, 2000).

As a result, we obtained the camera matrix of the ther-

mal camera K

therm

. The camera matrix K is generally

deﬁned as (Hartley and Zisserman, 2004):

K =





s p

0 α

0 0 1





(1)

This matrix contains the scale factors in x and y

direction α

and α

, the skew s and the x and y coor-

dinate p

and p

of the principal point.

In order to undistort the thermal image in ra-

dial and tangential direction, the method by Heikkila

(Heikkila and Silven, 1997) was applied. The RGB

images from the Kinect were not undistorted since

there were almost no distortions. However, the Kinect

was intrinsically calibrated to obtain the camera ma-

trix of the RGB sensor K

RGB

. For this calibration,

Skin Temperature Measurement based on Human Skeleton Extraction and Infra-red Thermography - An Application of Sensor Fusion

Methods in the Field of Physical Training

we applied the commonly used checkerboard target.

Both K

RGB

and K

therm

were needed for the extrinsic

calibration and the mapping, which are described in

the following sections.

3.3 Extrinsic Calibration

The calibration setup with the circle grid and the TFT

monitor that has been introduced in Section 3.2 was

used for the extrinsic calibration. Figure 2 shows the

view of both RGB and thermal sensor. We eluded

large rotations between the calibration target and the

sensors, so that elliptical shapes and consequent false

detections of the used circle detection algorithm could

be avoided. The centre points of the circles that were

detected in both RGB and undistorted thermal image

and their 3-D correspondences were used for the ex-

trinsic calibration that ﬁnally results in the rotation

matrix R and the translation vector t, see (Hartley and

Zisserman, 2004). R and t are given with respect to

the world coordinate system, which is located at the

RGB sensor of the Kinect. In other words, R and t de-

scribe the rotation and translation of the thermal sen-

sor with respect to the RGB sensor.

Figure 2: Calibration target in front of a TFT monitor, cap-

tured by the thermal camera (left) and the Kinect RGB cam-

era (right).

In the next step, the undistorted thermal image

could be mapped onto the RGB image by using the

obtained camera matrices K

RGB

and K

therm

, the rota-

tion matrix R and the translation vector t. This proce-

dure is described in the following section.

3.4 Mapping

The mapping of the undistorted thermal image onto

the RGB image comprised the back- projection of the

undistorted thermal image to a plane in the world with

a deﬁned distance µ from the sensor using K

therm

, and

the forward projection of the resulting world points to

the RGB image using K

RGB

The back projection can be generally described as

(Hartley and Zisserman, 2004):

X(µ) =



−1

(µx − p

)



, (2)

Figure 3: Thermal data was mapped to the RGB image.

The mapping result is shown in the present alpha-blended

image.

whereas X(µ) denotes the back-projected world

point at the distance µ. In our application, µ was set

to a value of 3 metres. M is the ﬁrst 3×3 sub-matrix

of the projection matrix P. p

is the fourth column of

P, which is generally deﬁned as follows:

P = K



R|t





M|p



(3)

When we back-projected the thermal image to the

world, K

therm

was used to calculate the projection ma-

trix P. Since the world coordinate system was located

at the RGB sensor, R and t that have been calcu-

lated during the extrinsic calibration were used for

this back-projection step. The back-projected points

from the thermal image X were then projected onto

the RGB image according to Equation 4, which de-

scribes the general forward projection. At this point,

P was calculated using K

RGB

as well as a unit matrix

for R and a zero vector for t.

x = PX (4)

The resulting coordinates x in the RGB image

were afterwards bi-linearly interpolated. The forward

projection was realised by means of a look-up-table.

The result of this mapping process is exemplary illus-

trated in Figure 3.

3.5 Skeleton Extraction

We localised the biceps brachii muscle by means of

the 2-D position of skeletal joints in the mapped im-

ages. For this purpose, we employed the skeleton ex-

traction algorithm that is provided by the Kinect ver-

sion 1.0 (Shotton et al., 2013). This algorithm calcu-

lates depth features from depth images and uses these

features to train a randomised decision forest. With

the help of this classiﬁer, both 3-D and 2-D positions

of overall 20 joints can be predicted. In our approach,

we used only the 2-D elbow and the shoulder joints

for the biceps brachii localisation on both left and

right upper arm.

VISAPP 2017 - International Conference on Computer Vision Theory and Applications

3.6 Skin Temperature Measurement

The biceps brachii temperature measurement using

shoulder and elbow joints is illustrated in Figure 4.

We discovered that the original shoulder joints 5

and 9 were not suitable to accurately localise the bi-

ceps brachii. Therefore, their original locations were

shifted by a parametrisable distance along the con-

necting, dashed line between the left and right shoul-

der in outward direction.

Figure 4: Left: Skin temperature measurement principle

using shifted shoulder joints 5’ and 9’ and elbow joints 6

and 10. Joint indices are according to the Kinect indexing.

Right: Thermal image with localised biceps brachii (black

line).

In addition to this modiﬁcation, we only used a

parametrisable percentage of the pixels lying on the

line connecting 5’ and 6 or 9’ and 10 respectively,

staring symmetrically from the centre of this line. In

this study, the percentage was set to 60 %, which re-

sults in a number of R and L temperature values for

the right and left biceps brachii respectively. Finally,

the skin temperature on the biceps brachii of the right

arm ϑ

(t) and the left arm ϑ

(t) were calculated by

determining the mean of the R and L temperature val-

ues of the pixels located on the black lines. These

pixel values are denoted as temp

(t) and temp

(t) re-

spectively, see Equation 5. t corresponds to the mea-

surement point.

(t) =

∑

r=1

temp

(t) (5a)

(t) =

∑

l=1

temp

(t) (5b)

4 EVALUATION

METHODOLOGY

In our experiments, we measured the biceps brachii

temperature of the left and the right arm for four per-

sons (P = 4) while they performed three sets of biceps

curls with their right arm. The weight of the dumb-

bell was 7 kg. The distance from the camera was ap-

proximately 3 meters. The measurement schedule is

presented in Table 1. At the beginning, each person

had a time of 15 minutes for acclimatisation. The ﬁrst

measurement was taken directly after this acclimati-

sation phase (t = 1). Then every person had to per-

form as many biceps curls as possible for every set.

Directly after each set, a measurement was performed

(t = 2, 3, 4). The pause between two consecutive sets

was 1 min. After the last set had been performed,

nine further measurements were taken, whereas the

time difference given in Table 1 is relative to the mea-

surement with index t=4. The measurements for both

arms were taken synchronously.

For evaluation, the measurements were equidis-

tantly displayed, which means that the distance does

not correspond to the time that elapsed between the

measurement points.

We used three ways to generate temperature pro-

ﬁles. The ﬁrst proﬁle exemplarily displays the abso-

lute temperature measured for the right arm ϑ

(t) and

left arm ϑ

(t) of one of the probands, see Figure 5.

For the second proﬁle, see Figure 6, we aimed at vi-

sualising the averaged relative changes ϑ

r,rel

(t) and

l,rel

(t) with respect to the starting temperature t

each arm, see Equations 6. The starting temperature

corresponds to the measurement time t = 1, i. e. after

acclimatisation. This type of representation was re-

quired for the evaluation of different probands’ data

to eliminate the inﬂuence of different environmental

temperatures for the different probands.

r,rel

(t) =

∑

p=1

(ϑ

(t) − ϑ

)) (6a)

l,rel

(t) =

∑

p=1

(ϑ

(t) − ϑ

)) (6b)

The third proﬁle illustrates the averaged difference

between the right and the left arm for each measure-

Table 1: Measurement schedule.

Measurement Point t Description

1 After acclimatisation

2 After ﬁrst set

3 After second set

4 After third set

5 After 1 min

6 After 2 min

7 After 3 min

8 After 4 min

9 After 5 min

10 After 10 min

11 After 15 min

12 After 20 min

13 After 25 min

Skin Temperature Measurement based on Human Skeleton Extraction and Infra-red Thermography - An Application of Sensor Fusion

Methods in the Field of Physical Training

ment time and is deﬁned as

diff

(t) =

∑

p=1

(ϑ

(t) − ϑ

(t)) . (7)

This proﬁle type allows to eliminate environmen-

tal changes during a proband’s training session and is

visualised in Figure 7.

5 RESULTS AND DISCUSSION

In this section, the obtained temperature proﬁles that

were described above are presented. The aim of these

analyses is to investigate the difference between the

temperature of the right biceps, which was active, and

the left biceps, which was passive. We discuss every

proﬁle and explain occurring effects.

Absolute Temperatures of a Single Person. Fig-

ure 5 compares the absolute temperatures of the bi-

ceps of the right and the biceps of the left arm. From

this ﬁgure, we can see that the temperature of the right

biceps increases more than the temperature of the left

biceps. What is interesting is that the maximal tem-

perature of the active biceps was reached only several

minutes after the training. These results are consistent

with the ﬁndings of Formenti et al. (Formenti et al.,

2013).

1 2 3 4

5 6

7 8 9 1011 1213

Measurement Point t

(t) and ϑ

(t) in

◦

Absolute Temperature

left

right

Figure 5: Absolute temperatures over time for one proband.

A possible explanation for this behaviour is that

the cardiovascular system reacts to the physical stress

by providing more blood to the active muscle. The

blood has core body temperature, which results in the

warming of the muscle. The adjacent tissue warms

up as well due to thermal conduction. This described

process needs time and results in the delay that is vis-

ible in the graph.

Temperatures Relative to the Starting Tempera-

ture of all Persons. To compare the temperature be-

haviour of the right and left biceps for more than one

proband, we examined the mean relative temperature,

see Figure 6. This type of evaluation allows the de-

termination of temperature changes with respect to a

start time. Taking the average of all probands, we ob-

tained results that are comparable to the previous pro-

ﬁle with one proband. Moreover, it can be seen that

the temperature of the right arm increases by approx-

imately 1.5

◦

C on average. The temperature of the

left arm decreases slightly at the beginning. An ex-

planation for this is that the blood ﬂow might even be

reduced in the passive arm.

1 2 3 4

5 6

7 8 9 1011 1213

−1

Measurement Point t

r,rel

(t) and ϑ

l,rel

(t) in

◦

Relative Temperature

left

right

Figure 6: Mean relative temperature with respect to start

temperature over time for all probands.

Temperature Difference Between Arms of all Per-

sons. In order to eliminate environmental inﬂuences

during one training session, we evaluated the mean

difference between the temperatures of the right and

the left arm. By doing this, eventual temperature

changes in the training room would not inﬂuence the

measurement, because this change would inﬂuence

both arms. Moreover, by employing this evaluation

method, we can determine the temperature difference

compared to the corresponding limb, i. e. the left arm,

which is passive. Figure 7 provides the according

temperature proﬁle.

The graph illustrates that, after the ﬁrst set, the

right arm has a higher temperature as the left arm.

Another ﬁnding to emerge from this graph is that, av-

eraged over all probands, the right arm is more than

◦

C warmer than the left arm at the maximum.

VISAPP 2017 - International Conference on Computer Vision Theory and Applications

1 2 3 4

5 6

7 8 9 1011 1213

0.5

Measurement Point t

diff

(t) in

◦

Difference Temperature

Figure 7: Mean temperature difference between right and

left arm over time for all probands.

6 CONCLUSIONS AND FUTURE

WORK

In this study, we presented a method to automati-

cally determine and evaluate skin temperatures. This

method is based on a sensor fusion of a thermal cam-

era and the Kinect. In order to fuse both sensor data,

we introduced a novel calibration procedure and de-

signed a special calibration target. The obtained re-

sults provide further support for the hypothesis that

the skin temperature increases during and after the

training. Moreover, we evaluated relative temperature

measurements, i. e. differences between active and

passive muscles, instead of absolute measurements.

This allows the elimination of environmental changes

in a training session.

Further research could investigate the inﬂuence of

subcutaneous fat tissue and clothing on the thermal

conduction. Moreover, future research might explore

skin temperature proﬁles of other muscles as well.

Another aspect in our future work will be the detec-

tion of skeleton joints directly on the thermal image,

which could considerably simplify the sensor system.

Continued efforts are needed to transfer our ap-

proach to other applications ﬁelds, such as medical

diagnostics, which can proﬁt from automatic temper-

ature measurements.

ACKNOWLEDGEMENTS

This project is funded by the European Social Fund

(ESF). We furthermore would like to express our

thanks to all the persons who contributed to this

project during the recordings.

REFERENCES

Bandeira, F., Moura, M. A. M. d., Souza, M. A. d., No-

hama, P., and Neves, E. B. (2012). Can thermography

aid in the diagnosis of muscle injuries in soccer ath-

letes? Revista Brasileira de Medicina Do Esporte,

18(4):246–251.

Bartuzi, P., Roman-Liu, D., and Wi

sniewski, T. (2012).

The inﬂuence of fatigue on muscle temperature. In-

ternational Journal of Occupational Safety and Er-

gonomics, 18(2):233–243.

Ellmauthaler, A., da Silva, E. A., Pagliari, C. L., Gois, J. N.,

and Neves, S. R. (2013). A novel iterative calibration

approach for thermal infrared cameras. In 2013 IEEE

International Conference on Image Processing, pages

2182–2186. IEEE.

Fern

andez-Cuevas, I., Marins, J. C. B., Lastras, J. A., Car-

mona, P. M. G., Cano, S. P., Garc

ıa-Concepci

on,

A., and Sillero-Quintana, M. (2015). Classiﬁca-

tion of factors inﬂuencing the use of infrared thermog-

raphy in humans: A review. Infrared Physics & Tech-

nology, 71:28–55.

Formenti, D., Ludwig, N., Gargano, M., Gondola, M.,

Dellerma, N., Caumo, A., and Alberti, G. (2013).

Thermal imaging of exercise-associated skin tempera-

ture changes in trained and untrained female subjects.

Annals of biomedical engineering, 41(4):863–871.

ohlich, M., Ludwig, O., Kraus, S., and Felder, H. (2014).

Changes in skin surface temperature during muscular

endurance indicated strain-an explorative study. In-

ternational Journal of Kinesiology & Sports Science,

2(3):23.

Gschwandtner, M., Kwitt, R., Uhl, A., and Pree, W. (2011).

Infrared camera calibration for dense depth map con-

struction. In Intelligent Vehicles Symposium (IV),

2011 IEEE, pages 857–862. IEEE.

Hartley, R. I. and Zisserman, A. (2004). Multiple View Ge-

ometry in Computer Vision. Cambridge University

Press, ISBN: 0521540518, second edition.

Heikkila, J. and Silven, O. (1997). A four-step camera cal-

ibration procedure with implicit image correction. In

Proceedings of the 1997 Conference on Computer Vi-

sion and Pattern Recognition (CVPR ’97), CVPR ’97,

pages 1106–, Washington, DC, USA. IEEE Computer

Society.

Hildebrandt, C., Zeilberger, K., Ring, E. F. J., and Raschner,

C. (2012). The application of medical infrared ther-

mography in sports medicine. INTECH Open Access

Publisher.

Lahiri, B., Bagavathiappan, S., Jayakumar, T., and Philip,

J. (2012). Medical applications of infrared thermog-

raphy: a review. Infrared Physics & Technology,

55(4):221–235.

Ludwig, N., Formenti, D., Trecroci, A., Gargano, M., and

Alberti, G. (2014). Comparison of image analysis

methods in skin temperature measurements during

physical exercise.

Nakagawa, W., Matsumoto, K., de Sorbier, F., Sugimoto,

M., Saito, H., Senda, S., Shibata, T., and Iketani, A.

(2014). Visualization of temperature change using

Skin Temperature Measurement based on Human Skeleton Extraction and Infra-red Thermography - An Application of Sensor Fusion

Methods in the Field of Physical Training

RGB-D camera and thermal camera. In Workshop at

the European Conference on Computer Vision, pages

386–400. Springer.

Neves, E. B., Vilac¸a-Alves, J., Krueger, E., and Reis, V. M.

(2014). Changes in skin temperature during muscular

work: a pilot study. Pan American Journal of Medical

Thermology, 1(1):11–15.

Rangel, J., Soldan, S., and Kroll, A. (2014). 3D thermal

imaging: Fusion of thermography and depth cameras.

In International Conference on Quantitative InfraRed

Thermography.

Richter, J., Wiede, C., Apitzsch, A., Nitzsche, N., L

osch, C.,

Weigert, M., Kronfeld, T., Weisleder, S., and Hirtz,

G. (2016). Assisted Motion Control in Therapy En-

vironments Using Smart Sensor Technology: Chal-

lenges and Opportunities. In Proceedings of Zukunft

Lebensr

aume Kongress 2016, pages 90–102.

Ring, E. and Ammer, K. (2012). Infrared thermal imaging

in medicine. Physiological measurement, 33(3):R33.

Ring, E. and Ammer, K. (2015). The technique of infrared

imaging in medicine. In Infrared Imaging. IOP Pub-

lishing.

Shotton, J., Girshick, R., Fitzgibbon, A., Sharp, T., Cook,

M., Finocchio, M., Moore, R., Kohli, P., Criminisi,

A., Kipman, A., et al. (2013). Efﬁcient human pose

estimation from single depth images. Pattern Analy-

sis and Machine Intelligence, IEEE Transactions on,

35(12):2821–2840.

Suzuki, S. and Abe, K. (1985). Topological structural anal-

ysis of digitized binary images by border following.

Computer Vision, Graphics, and Image Processing,

30(1):32–46.

Thomanek, J., Ritter, M., Lietz, H., and Wanielik, G.

(2011). Comparing Visual Data Fusion Techniques

Using FIR and Visible Light Sensors to Improve

Pedestrian Detection. pages 119–125. IEEE.

Vidas, S., Moghadam, P., and Bosse, M. (2013). 3d thermal

mapping of building interiors using an RGB-D and

thermal camera. In Robotics and Automation (ICRA),

2013 IEEE International Conference on, pages 2311–

2318. IEEE.

Zhang, Z. (2000). A ﬂexible new technique for camera cal-

ibration. IEEE Transactions on pattern analysis and

machine intelligence, 22(11):1330–1334.

VISAPP 2017 - International Conference on Computer Vision Theory and Applications