Application of LSD-SLAM for Visualization Temperature in Wide-area

Environment

Masahiro Yamaguchi

, Hideo Saito

and Shoji Yachida

Graduate School of Science and Technology, Keio University, 3-14-1 Hiyoshi Kohoku-ku, Yokohama, Kanagawa, Japan

NEC Corporation, 1753 Shimonumabe Nakahara-ku, Kawasaki, Kanagawa, Japan

{yamaguchi, saito}@hvrl.ics.keio.ac.jp, s-yachida@bk.jp.nec.com

Keywords:

Thermal Camera, SLAM, RGB Camera.

Abstract:

In this paper, we propose a method to generate a three-dimensional (3D) thermal map by overlaying thermal

images onto a 3D surface reconstructed by a monocular RGB camera. In this method, we capture the target

scene moving both an RGB camera and a thermal camera, which are mounted on the same zig. From the

RGB image sequence, we reconstruct 3D structures of the scene by using Large-Scale Direct Monocular

Simultaneous Localization and Mapping (LSD-SLAM), on which temperature distribution captured by the

thermal camera is overlaid, thus generate a 3D thermal map. The geometrical relationship between those

cameras is calibrated beforehand by using a calibration board that can be detected by both cameras. Since we

do not use depth cameras such as Kinect, the depth of the target scene is not limited by the measurement range

of the depth camera; any depth range can be captured. To demonstrating this technique, we show synthesized

3D thermal maps for both indoor and outdoor scenes.

1 INTRODUCTION

Thermal cameras can be used to detect a gas leaks,

ﬁres, abnormalities in electronic apparatus and so on,

many of which cannot be detected with visible light.

However, thermal camera have some restrictions. For

example, the angle of a thermal camera is narrower

than the angle of RGB cameras because of the ma-

terial composing the thermal camera lens. Obtain-

ing thermal images with a wide area requires multiple

ﬁxed thermal cameras or moving the thermal camera.

However, it is difﬁcult to extract local feature points

from thermal images, so we develop a method of ob-

taining wide-range thermal images by using a method

like image stitching. It is also difﬁcult to recognize

the shapes of three-dimensional (3D) structures if we

only use thermal images. For instance, we cannot un-

derstand where the thermal image is taken when we

acquire a thermal image of large machines from a

close distance or of a line of similar objects. Thus, in

related works, other cameras are used in conjunction

with thermal cameras to compensate for the short-

comings of thermal cameras, allowing us to more eas-

ily understand the acquired thermal information. Us-

ing RGB-D cameras is one of those methods. Ther-

mal information is superimposed on a 3D structure in

that method. Matsumoto makes an augmented real-

ity (AR) application system that can be visualized the

thermal distribution on the environment (Matsumoto

et al., 2015). They use kinect for RGB-D camera and

thermal camera. In the off-line phase, they can ob-

tain a 3D model of the environment, and in the on-

line phase, they track the camera in order to observe

the thermal map for a wide area from any viewpoint.

However, it is difﬁcult to obtain depth value in out-

door environment because sunlight affects the depth

sensor and objects are often too far away to get depth

value.

In this paper, we propose a method to visualize

thermal information as a 3D structure and superim-

pose thermal information on arbitrary RGB images.

Using a low cost depth sensor would prevent us from

experimenting in outdoor environment, so we use a

monocular RGB camera and a thermal camera in our

works. We create 3D models of the environment with

thermal information to superimpose thermal informa-

tion on arbitrary RGB images. The 3D model is made

by Simultaneous Localization and Mapping (SLAM),

which is one of the methods to reconstruct an environ-

ment as a 3D point cloud. We can track camera po-

sitions and reconstruct 3D points simultaneously with

only a monocular RGB camera by using SLAM. We

add thermal information to the 3D points cloud recon-

structed by SLAM. After that, we project the point

216

Yamaguchi M., Saito H. and Yachida S.

Application of LSD-SLAM for Visualization Temperature in Wide-area Environment.

DOI: 10.5220/0006153402160223

In Proceedings of the 12th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2017), pages 216-223

ISBN: 978-989-758-225-7

cloud onto arbitrary RGB images and create a mesh

by Delaunay division. As a result, our system can be

used to image both indoor and outdoor scenes because

we use a monocular RGB camera.

This paper is structured as follow. In Section 2,

we discuss related works. We provide details of our

proposed system in Section 3 and show the results of

our experiment in Section 4. Finally, we describe our

conclusions and suggest directions for future work in

Section 5.

2 RELATED WORKS

Our goal is to visualize temperature of the environ-

ment as a 3D structure by using a monocular RGB

camera. However, most works use RGB-D camera to

achieve similar visualization. Thus, we discuss such

methods as related works. We then introduce some

SLAM methods to reconstruct 3D point clouds by us-

ing a monocular RGB camera.

2.1 Making a Thermal Map

One method of visualizing thermal images over a

wide area is to obtain the 3D structure of the envi-

ronment and render temperatures onto it. (Borrmann

et al., 2012) make 3D thermal map by using a robot,

a razer scanner, and a thermal camera . The robot

repeatedly moves and stops to take a razer scan and

make a 3D thermal map. The outputs are both ther-

mal and RGB point cloud. However, the robot moves

on wheels so they cannot use it to obtain data with

poor footing and it cannot climb over steps. (Ham

and Golparvar-Fard, 2013) use a camera that captures

both RGB and thermal frames and creates thermal

map by using Structure From Motion and Multi-View

Stereo images. However, this method requires a gen-

eral processing unit (GPU) to carry out the calcula-

tions. (Vidas et al., 2013) and (Nakagawaet al., 2014)

make thermal maps by using an RGB-D camera and a

thermal camera. They move ﬁx the thermal and RGB-

D cameras to each other and move them continuously

to make a 3D thermal map. This equipment is hand

held and low cost. However, the RGB-D cameras

used in such works are limited in that they cannot be

used in outdoor environment. Low-cost RGB-D cam-

eras cannot obtain depth values if objects are far from

the thermal camera, and sunlight also affects the abil-

ity of RGB-D cameras to obtain depth values. Thus,

in our work, we overcome this limitation by propos-

ing a method that uses RGB cameras so that 3D ther-

mal maps can be obtained in outdoor environment.

2.2 Monocular RGB SLAM

We use a monocular RGB camera together with a

thermal camera in order to make a experiment in out-

door environment. Thus, the RGB-D camera method

of reconstructing 3D structure cannot be used. In-

stead, we use SLAM for reconstruction. SLAM is one

method of obtaining 3D structures by using a monoc-

ular RGB camera. We can get camera position and 3D

points simultaneously by using SLAM. One SLAM

method is Parallel Tracking and Mapping (PTAM),

which was proposed by (Klein and Murray, 2007).

PTAM can estimate camera pose and translation in

real time because PTAM uses parallel threads to do

mapping and tracking. While PTAM uses feature

points to estimate camera pose and translation, it can-

not use them to estimate where the camera is within a

world coordinate system. Thus, PTAM cannot detect

large loop closure, so it is not capable of large-scale

3D reconstructions. (Mur-Artal et al., 2015) propose

Oriented FAST and Rotated BRIEF SLAM (ORB-

SLAM) which refers to the same method as PTAM. In

their work, feature points are used in tracking, map-

ping, position matching, and loop detection. Thus,

this work can estimate camera pose and translation

on a large scale. However, the number of points in the

map is small because they use ORB-feature points, so

the number of points that can be reconstructed within

a frame is small.

(Engel et al., 2014) proposed LSD-SLAM, which

estimates camera pose and translation not by using

feature points but by using local distribution matching

within a frame. Point clouds made by LSD-SLAM

are more dense than point clouds made by PTAM or

ORB-SLAM because LSD-SLAM retains the pixel

values of each frame. Figure 1 shows an example

of 3D reconstructions by PTAM and by LSD-SLAM.

Our goal is to visualize temperature as a 3D structure,

so density mapping suits our purpose. We thus use

LSD-SLAM to make our thermal maps.

(a) PTAM (b) LSD-SLAM

Figure 1: Examples of reconstruction by using SLAM.

Application of LSD-SLAM for Visualization Temperature in Wide-area Environment

217

3 PROPOSED SYSTEM

The purpose of our system is to visualize thermal in-

formation as a 3D model, superimposing thermal in-

formation on the RGB images. Our system consists of

three stages: camera calibration, 3D reconstruction,

and superimposition of the images. First, we ﬁx the

RGB and thermal cameras and calibrate them to ob-

tain the intrinsic and extrinsic parameters. After that,

we move the equipment to obtain a video sequence.

We then reconstruct a 3D point cloud by using LSD-

SLAM and add the thermal information to each point

cloud. Finally, we superimpose thermal information

on RGB images and create a mesh by Delaunay divi-

sion.

Thermal images have less features of brightness

than RGB images, and temperature changes more

smoothly than RGB values. Thus, commonly used

key point descriptors like SIFT (Lowe, 1999) or

SURF (Bay et al., 2006) fail to obtain feature points.

This makes it difﬁcult to track camera motion or re-

construct structure from motion. The direct method

(Engel et al., 2013) that do not use feature points such

as LSD-SLAM, are also not able to track camera mo-

tion because thermal images do not produce enough

features of brightness. Figure 3 shows the difference

between semi-dense map for the same scene acquired

by each camera. The semi-dense map for the thermal

camera produces less depth value than the map from

the RGB camera, and tracking is lost if only the ther-

mal camera is used. Thus, using another camera is

a better way to visualize thermal information as 3D

structures.

3.1 Camera Calibration

Because the temperature of calibration boards is con-

stant, we cannot calibrate the thermal cameras in the

same way that we calibrate RGB cameras. Thus, we

must develop a calibration board that can be captured

by both RGB cameras and thermal cameras. (Prakash

et al., 2006) heated a calibration board with a ﬂood

lamp and based on the emissivity difference between

the black and white regions, theycan detect the check-

ered pattern. We developed a checkered calibration

board that creates differences in temperature within

the pattern. We use an electric carpet and thermal in-

sulation material to generate the temperature differ-

ence, allowing calibration images to be obtained like

those shown in Figure 2. Based on this calibration

board, we can calibrate both the thermal and RGB

cameras. We use the method of (Zhang, 2000) to cal-

culate the internal parameter.

3.2 Reconstruct the Thermal 3D Point

Cloud

One of the methods of reconstructing 3D models by

using a monocular camera is SLAM. This method

tracks camera position and rotation and maps 3D

points that are used to obtain the scene. In our works,

we use LSD-SLAM which is one of method of SLAM

and visualize thermal information as 3D structures by

using it. Figure 4 shows a ﬂow chart of the recon-

struction part of our system.

3.2.1 The Features of LSD-SLAM

Methods like PTAM and ORB-SLAM use a monoc-

ular RGB camera and detect feature points from im-

ages. In such methods, camera pose and translation

are estimated based on feature points, but only fea-

ture points are used to make the resulting map. Thus,

the 3D point clouds are very sparse. However, LSD-

SLAM uses the pixel values in the images to estimate

camera pose and translation, which results in more

robust estimations and denser 3D point clouds than in

previous works.

3.2.2 Process of LSD-SLAM

Camera pose and translation are estimated by com-

parison of the pixel values between the input frame

and the key frame. In LSD SLAM, a semi-dense map

of each input frame and its camera pose and posi-

tion are computed by matching with a key frame and

a 3D map recovered until the previous frame. The

semi-dense map is integrated with the 3D map. If the

matching score is less than a pre-deﬁned threshold,

the key frame is replaced with the input frame. In the

proposed method, temperature image is also captured

when the key frame is replaced with the input frame,

so that the temperature is also added to the 3D map.

3.2.3 Map Constitution

The factors contained in the key frame constitute the

map. Each key frame contains the pixel values of the

input images, the depth map, and dispersion of the

depth map. In this work, some pixel in depth map

do not retain depth value. The map only keeps the

depth information that satisﬁes a given threshold. The

threshold is based on the values of the peripheral pix-

els. This results in a semi-dense map. We can use

a CPU because semi-dense map calculation requires

less computation power.

VISAPP 2017 - International Conference on Computer Vision Theory and Applications

218

(a) (b) (c)

Figure 2: (a)Checker image taken by RGB camera. (b) Checker image taken by thermal camera. (c) Checker detection pattern

from thermal camera.

(a) Semi-dense map for

RGB camera

(b) Semi-dense map for

thermal camera

Figure 3: Less features are detected in the thermal image

compared to the RGB image.

3.2.4 Tracking a New Frame

is a vector representing translation from frame i to

frame j.

E(ξ

) =

∑

(P) − I

(w(P, D

(P), ξ

))

(1)

Equation 1 compares the pixel value at point P in

the key frame and the pixel value at point P in frame

j, which has been translated and is calculated by esti-

mated translation value ξ

and depth map in frame j.

Equation 1 is minimized by using the Gauss-Newton

method and we estimate camera pose and tracking.

3.2.5 Evaluation of Key Frame

When the camera is moved enough, the key frame is

uploaded. Whether the frame is a key frame is then

evaluated based on whether it satisﬁes the threshold

requirements. If the threshold is satisﬁed, the depth

map is initialized and the map is renewed. If the

threshold is not satisﬁed, the current key frame is up-

dated by adding depth information to the semi-dense

map of the current key frame.

3.2.6 Map Optimization

Maps created using monocular SLAM do not have

an inherent scale, which makes it difﬁcult to create a

large-scale map. LSD-SLAM, however, has an inher-

ent scale based on equation 2. Equation 2 optimizes

the depth, using equation 1.

E(ξ

) =

∑

x∈Ω

((I

(P) − I

′

))

+ ([P

′

]

− D

′

))

)

(2)

P’ indicates a warped point. After the key frame

is added to the map, frames that are candidates for

loop closure are collected, and suitable candidates are

detected by using an appearance-based mapping algo-

rithm (Glover et al., 2012).

3.2.7 Reprojection of the Point Cloud onto

Thermal Images

Semi-dense maps hold depth data that satisﬁes the

threshold set by the brightness of the surrounding fea-

tures. Because of this threshold, LSD-SLAM requires

only a CPU. Pixels that have color possess depth

value. Depth is shown in gradations of blue, green,

and red, where red is closer and blue is farther away.

The semi-dense map possessed by the key frame is

updated when the key frame is renewed. In this work,

the semi-dense map is reprojected onto the thermal

image that is taken when the key frame is renewed.

Thus, each pixel also contains thermal information.

The semi-dense map at pixel i is warped to the ther-

mal image by equation 3. I represents the semi-dense

map and I

represents the thermal image.

(i) = I(π(KTπ

−1

(i, d))) (3)

K is the intrinsic matrix of the thermal camera and

T described the relative positional relation between

the thermal camera and the RGB camera. π

−1

(i, d) =

−1

i is a conversion function, meaning that pixel i

Application of LSD-SLAM for Visualization Temperature in Wide-area Environment

219

Figure 4: Flow chart of visualize temperature as 3D structures. KF in this ﬁgure means key frame.

is converted to a 3D point according to depth d. In

Figure 5, the black part shows the reprojection from

the semi-dense map; each pixel possesses the thermal

information of the projectable part. In the 3D map

reconstruction step, each 3D points contains thermal

information. Thus, we can obtain a 3D thermal map.

(a) Semi-dense map for

LSD-SLAM

(b) Reprojection

semi-dense map

Figure 5: Example of reprojection.

3.3 Superimpose of Thermal

Information onto RGB Images

Normally, we cannot obtain thermal information cor-

responding to each pixel of an RGB camera image be-

cause the optical axis is different between RGB cam-

eras and thermal cameras. Our system superimposes

temperature on the RGB images by using a corre-

sponding point cloud that is reconstructed by using

LSD-SLAM from both the RGB and thermal images.

First, the point cloud is reprojected to the RGB im-

ages by using the RGB cameras extrinsic parameter of

each frame, which is obtained by using LSD-SLAM.

Then, 2D Delaunay division is used to divide each

point into a triangle region, as shown in Figure 6.

Next, the triangles obtained by Delaunay division are

transferred to the thermal image by using translation

and rotation obtained from the camera calibration of

each camera. We obtain an afﬁne matrix from three

points by constructing a triangle in order to transfer

the triangle patch from the RGB image to the thermal

image. Finally, the triangle patches are transferred

from the thermal image back to the RGB image, and

using bilinear interpolation, we obtain the ﬁnal result.

Figure 6: Delaunay division.

4 EXPERIMENT

4.1 Experiment Environment

In this experiment, we ﬁxed a monocular RGB cam-

era and a thermal camera together into one hand-held

device. This device can be moved to capture video

of the environment. We used an HD PRO C920R in

Logicool (640 × 480 pixels resolution, 30fps), and

Optris PI640 (640 × 480 pixels resolution, 30fps).

Figure 8 shows our camera system.

VISAPP 2017 - International Conference on Computer Vision Theory and Applications

220

Example images

Visualization temperature as 3D model

Figure 7: Visualization result for an image of an interior environment. Upper left shows an example RGB image of the scene

and bottom left shows an example thermal image of the scene. The right panel shows the result of temperature visualization

as a 3D model. Temperature is shown in gradation of blue, green, and red, where red is higher and blue is lower.

Figure 8: Capture system composed of the HD PRO C920R

and Optris PI640 thermal cameras.

4.2 Temperature Visualization for the

3D Structure

We show results of temperature visualization as 3D

structures for both indoor and outdoor scenes. The

temperature of the plastic bottle in Figure 7 is lower

than the average image temperature while that of the

monitor is higher in the rendering results for the in-

door environmenyt. In Figure 9, the car in the middle

has the highest temperature in the outdoor image. We

can those determine that the car in the middle has just

been used, which cannot be judged using only visi-

ble light. We can also visualize the temperature of 3D

structures in outdoor environments like the parking

area shown.

However, there are some points for which thermal

information cannot be obtained. One of the reasons

for this may be that not every point reconstructed by

LSD-SLAM can be projected to the thermal camera

image because the imaging range of the thermal cam-

era and the RGB camera do not match. Thus, we re-

move such points from the points used to render the

image in the next environment.

Our system is applied to a static environment, so

good results cannot be obtained if the object is mov-

ing or the temperature is changing. Let us suppose

that our system is used for outdoor inspection work in

a wide area.

4.3 Creating Mesh

We discuss the result of temperature distribution ren-

dered onto RGB images. We can observe that some

regions are not correctly aligned with the correspond-

ing areas in RGB images in ﬁgure 10, especially the

plastic bottle and the display. This is because the scale

is undeﬁned for the point cloud that is reconstructed

by SLAM with the monocular RGB camera. In our

work, the point cloud obtained by SLAM is superim-

posed on the thermal images according to the rotation

and translation between both cameras. Although the

scale of the translation between both cameras is es-

timated based on the scale of the calibration board,

the point clouds scale is not able to obtained by the

monocular SLAM. We thus cannot align the position

accurately. Because of such uncertainly of the scale

of the point cloud reconstructed by LSD-SLAM, the

alignment between RGB images and temperature dis-

tribution in ﬁgure 10 is not correctly performed.

In order to correct the wrong alignment, we ﬁ-

Application of LSD-SLAM for Visualization Temperature in Wide-area Environment

221

Example images

Visualization temperature as 3D model

Figure 9: Visualization result for an image from an outdoor environment. Upper left shows an example RGB image of

the scene and bottom left shows an example thermal image of the scene. The right panel shows the result of temperature

visualization as a 3D model. Temperature is shown in gradation of blue, green, and red, where red is higher and blue is lower.

(a) Result of temperature

rendering

(b) Clicking

corresponding points

Figure 10: Example of temperature rendering.

nally manually aligned them by clicking on the cor-

responding points between RGB image and rendered

temperature distribution as shown in ﬁgure 10 (b). We

compute a homography between RGB image and the

rendered temperature distribution, because they can

be correctly aligned by a homography,which can rep-

resent scale difference in 3D point cloud. As a re-

sult, we can make a successful alignment as shown in

ﬁgure 11. We can conﬁrm that enough correspond-

ing is obtained by simple methods such as using the

homography because we have geometrically accurate

correspondence between RGB and thermal data by

3D structure. We plan to solve this uncertainty of the

scale to obtain more correct 3D thermal maps.

When we render the temperature distribution, we

ﬁrst use thermal frames corresponding to the key

frames of LSD SLAM for rendering of temperature

on RGB images. If a triangle in the Delaunay division

is out of the reference thermal image, we refer other

frames to obtain thermal distribution in the triangle.

5 CONCLUSION & FUTURE

WORK

In this paper, we visualize thermal information as 3D

structures and obtain experimental images from both

indoor and outdoor scenes. We also superimposed

thermal information onto RGB images by reproject-

ing the point cloud onto RGB images. As part of the

visualization system, we ﬁxed an RGB camera and

a thermal camera together as a hand-held device and

calibrated them by using a special calibration board

that can detect a checkered pattern with temperature

differences using both an RGB camera and a ther-

mal camera. We can thus obtain intrinsic parameters

of both cameras and their relative positional relation.

After that, the 3D point cloud is reconstructed by us-

ing LSD-SLAM based on image sequences taken by

the RGB camera and reprojected to the thermal im-

age. Each point is updated with thermal information

about the projectable parts. In the case of superim-

posing thermal information onto the RGB images, the

point cloud is reprojected onto the RGB image and

each point is divided into Delaunays triangles, which

are used to create a mesh framework. In our work,

visualization of outdoor temperatures as a 3D model

is possible. This is not possible by using methods de-

scribed in previous work that use kinect. We plan to

expand our work by using drones to acquire images

and make large-scale 3D thermal maps.

VISAPP 2017 - International Conference on Computer Vision Theory and Applications

222

indoor 100

frame

outdoor 100

frame

indoor 400

frame

outdoor 600

frame

indoor 1300

frame

outdoor 900

frame

Figure 11: Result of rendering temperature. Each caption shows the frame number. Top shows the result for an indoor scenes

and bottom shows the result for an outdoor scenes. Temperature is shown in gradation of blue, green, and red, where red is

higher and blue is lower.

REFERENCES

Bay, H., Tuytelaars, T., and Van Gool, L. (2006). Surf:

Speeded up robust features. In European conference

on computer vision, pages 404–417. Springer.

Borrmann, D., N¨uchter, A., Dakulovi´c, M., Maurovi´c, I.,

Petrovi´c, I., Osmankovi´c, D., and Velagi´c, J. (2012).

The project thermalmapper–thermal 3d mapping of

indoor environments for saving energy. IFAC Pro-

ceedings Volumes, 45(22):31–38.

Engel, J., Sch¨ops, T., and Cremers, D. (2014). Lsd-slam:

Large-scale direct monocular slam. In European Con-

ference on Computer Vision, pages 834–849. Springer.

Engel, J., Sturm, J., and Cremers, D. (2013). Semi-dense

visual odometry for a monocular camera. In Proceed-

ings of the IEEE international conference on com-

puter vision, pages 1449–1456.

Glover, A., Maddern, W., Warren, M., Reid, S., Milford, M.,

and Wyeth, G. (2012). Openfabmap: An open source

toolbox for appearance-based loop closure detection.

In Robotics and Automation (ICRA), 2012 IEEE Inter-

national Conference on, pages 4730–4735. IEEE.

Ham, Y. and Golparvar-Fard, M. (2013). An automated

vision-based method for rapid 3d energy performance

modeling of existing buildings using thermal and dig-

ital imagery. Advanced Engineering Informatics,

27(3):395–409.

Klein, G. and Murray, D. (2007). Parallel tracking and map-

ping for small ar workspaces. In Mixed and Aug-

mented Reality, 2007. ISMAR 2007. 6th IEEE and

ACM International Symposium on, pages 225–234.

IEEE.

Lowe, D. G. (1999). Object recognition from local scale-

invariant features. In Computer vision, 1999. The pro-

ceedings of the seventh IEEE international conference

on, volume 2, pages 1150–1157. Ieee.

Matsumoto, K., Nakagawa, W., Saito, H., Sugimoto, M.,

Shibata, T., and Yachida, S. (2015). Ar visualiza-

tion of thermal 3d model by hand-held cameras. In

SciTePress.

Mur-Artal, R., Montiel, J., and Tard´os, J. D. (2015). Orb-

slam: a versatile and accurate monocular slam system.

IEEE Transactions on Robotics, 31(5):1147–1163.

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

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

(2014). Visualization of temperature change using

rgb-d camera and thermal camera. In Workshop at

the European Conference on Computer Vision, pages

386–400. Springer.

Prakash, S., Lee, P. Y., and Caelli, T. (2006). 3d mapping

of surface temperature using thermal stereo. In 2006

9th International Conference on Control, Automation,

Robotics and Vision.

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

mal 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.

Application of LSD-SLAM for Visualization Temperature in Wide-area Environment

223