Detection of Control Points for UAV-Multispectral Sensed Data

Registration through the Combining of Feature Descriptors

Jocival Dantas Dias Junior, Andr

e Ricardo Backes and Maur

ıcio Cunha Escarpinati

Faculty of Computing, Federal University of Uberl

andia, Uberl

andia/MG, Brazil

Keywords:

Image Registration, Unmanned Aerial Vehicle, Multispectral Image, Feature Descriptors.

Abstract:

The popularization of the Unmanned Aerial Vehicle (UAV) and the development of new sensors has enabled

the acquisition and use of multispectral and hyperspectral images in precision agriculture. However, perfor-

ming the image registration process is a complex task due to the lack of image characteristics among the

various spectra and the distortions created by the use of the UAV during the acquisition process. Therefore,

the objective of this work is to evaluate different techniques for obtaining control points in multispectral ima-

ges of soybean plantations obtained by UAVs and to investigate if combining features obtained by different

techniques generates better results than when used individually. In this work Were evaluated 3 different fea-

ture detection algorithms (KAZE, MEF and BRISK) and their combinations. Results shown that the KAZE

technique, achieve better results.

1 INTRODUCTION

It is becoming increasingly more common to see ima-

ging technologies used to aid agriculture in terms of

providing precision based tasks, either to estimating

crop growth or when identifying characteristics of

agronomic interest (Sankaran et al., 2015). In this sce-

nario, the use of unmanned aerial vehicles (UAVs) has

gained more and more space due to the reduction of

operational costs regarding the use of such technology

(Zecha et al., 2013). According to the latest economic

report by the Association of Unmanned Aerial Vehi-

cles International, precision agriculture occupies the

largest portion of the potential worldwide market for

UAVs. (AUVSI, 2013).

Sensors represent a fundamental part of the ima-

ging process, a variety of sensors are being used to

scan plants for health problems, record growth rates

and hydration, and locate disease outbreaks. The ﬁrst

UAVs used regular comercial cameras that operated

in the red, green and blue bands (RGB) and / or in re-

gions near the infrared (Hunt et al., 2010). The newly

developed sensors offered the UAVs the possibility

of obtaining multispectral and hyperspectral images

(Berni et al., 2009).

However, despite the growing use of UAVs to

obtain low and medium altitude images (100 to 400

m), the techniques of image processing used requires

specialized software. The reason for this is that con-

ventional methods applied to remote sensing image

processing are not applied to images obtained by UAV

as these methods have been developed to perform data

processing on more stable images with a much larger

spatial extent than the images obtained by UAVs (So-

ares et al., 2018).

Using UAVs for image acquisition in precision

agriculture requires hundreds and in some cases thou-

sands of overlapping images to cover an area. After

acquiring aerial images, it is necessary to perform the

registration process of the acquired images, in order

to extract agronomic characteristics. In RGB ima-

ges this process presents some difﬁculties that are ea-

sily identiﬁed and resolved, such as changes in lig-

hting, rotations and changes in scale from unforeseen

events along the UAV path. However, in addition to

the previously mentioned problems for RGB images,

we have that most multispectral cameras use different

physical sensors to obtain different spectra, which

causes a spatial misalignment between the spectra due

to their physical displacement. The variation of the

analyzed spectrum also leads to a loss of characteris-

tics between the bands which hinders the process of

detection of common characteristics between bands.

According to (Banerjee et al., 2018) the registra-

tion of two channels is achieved by inferring the ne-

cessary transformations from a set of control-point

correspondences that pair identical points in the scene

on each of the two images. In general, the greater the

444

Dias Junior, J., Backes, A. and Escarpinati, M.

Detection of Control Points for UAV-Multispectral Sensed Data Registration through the Combining of Feature Descriptors.

DOI: 10.5220/0007580204440451

In Proceedings of the 14th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2019), pages 444-451

ISBN: 978-989-758-354-4

number of identiﬁed points between the images, the

better the alignment of the channels. The traditional

approach for multispectral image registration is to de-

signate one channel as the target channel and register

all other image channels to the target. There is cur-

rently no comparative assessment of the best possible

way to perform such a registering procedure.

A framework for the registration of multispectral

images in spectral complex environments within the

temporal and spectral order is proposed in (Banerjee

et al., 2018). The descriptors Harris-Stephens Featu-

res (HSF), Min Eigen Features (MEF), Scale Invari-

ant Feature Transformation (SIFT), Speeded-Up Ro-

bust Features (SURF), Binary Robust Invariant Scala-

ble Keypoints (BRISK) and Features from Accelera-

ted Segment Test (FAST) were evaluated for this pro-

blem. The registration of these images in the spectral

order obtained a superior result to the temporal re-

gistration, where the best result was obtained by the

SURF method. However, the authors state that the use

of other descriptors can signiﬁcantly improve results.

In (Yasir, 2018), an automatic framework was pro-

posed for the registration of multispectral images that

deﬁne the target channel based on the assumptions

that a minimum number of control-points correspon-

dences between two channels is needed to ensure low-

error registration, and a greater number of such cor-

respondences generally results in higher registration

performance. Basically, this work consists of analy-

zing all spectra in pairs and identifying the best way

to perform the registration so that the steps for the re-

gistration of all bands have on average the largest set

of control points possible.

In (Junior et al., 2018) the authors performed a

comparative analysis between the main descriptors of

local characteristics in the context of multispectral re-

gistration of images obtained by UAVs. In this work,

Harris-Stephens Features (HSF), Min Eigen Features

(MEF), KAZE Features (KAZE), Speeded-Up Ro-

bust Features (SURF), Binary Robust Invariant Sca-

lable Keypoints (BRISK) and Features from Accele-

rated Segment Test (FAST) were analyzed. The aut-

hors concluded that algorithms that use corner featu-

res provide better alignment of multispectral images

of crops, and MEF was considered the best algorithm

for this process.

In (Faria, 2018) a combination approach of diffe-

rent descriptors is proposed for improving the classi-

ﬁcation of the interesting cells. Results showed that

the union of different features descriptors generates a

better classiﬁcation of cells than the individual appli-

cation of the same ones.

In this paper we propose the aplication of the ap-

proach proposed by (Faria, 2018) for the registration

process of aerial images. Our objective is to investi-

gate if the union of descriptors can be used with the

framework proposed by (Yasir, 2018) in order to im-

prove multispectral image registration. To accomplish

this task we used the following set of descriptors: Bi-

nary Robust Invariant Scalable Keypoints (BRISK),

Min Eigen Features (MEF), Kaze Features (KAZE)

and a combination between MEF and BRISK, MEF

and KAZE, BRISK and KAZE, because, according

to (Junior et al., 2018), these algorithms obtained on

average a superior result in the registration of multis-

pectral crop images.

The authors of the present paper conducted ex-

periments on aerial images of soybean plantations.

These images were chosen due to their peculiar cha-

racteristics that hinder the registration process. Soy-

bean images have a very similar texture and do not

usually contain much information (e.g. roads, lines,

trees) that can be used as control points for later alig-

nment.

The remainder of this paper is organized as fol-

lows. In section 2, the authors describe the dataset

and the main concepts used for the development of

this work. Section 3 presents the experiments and re-

sults obtained. Section 4 presents conclusions, limi-

tations and future work.

2 METHODS

In this section, the authors of the present paper dis-

cuss the datasets used in this work and the characte-

ristics observed during the acquisition of these ima-

ges. A description of the techniques of extraction and

detection of features is carried out. Finally, a brief ex-

planation is presented of the framework used to ﬁnd

the best way to register each dataset.

2.1 Dataset

Three datasets were used, in all cases the data-

sets are from soybean plantations with a size of

1280 × 960, a resolution of 96 dpi and average 75%

overlap between the images. The channels pre-

sent in the datasets are, respectively, blue, green,

red, near-IR (NIR) and red-edge (REDEG). The da-

tasets were obtained in different soybean plantati-

ons located at the following decimal coordinates (-

20.379918, -46.242159), (-20.448603, -46.308684)

and (-18.730114, -48.772294) respectively. Images

from each dataset were obtained on a single ﬂight wit-

hout any kind of pre-processing. The images were

obtained by a MicaSense Red-Edge (see Figure 1)

(MicaSense Inc. Seattle, WA, USA) camera coupled

Detection of Control Points for UAV-Multispectral Sensed Data Registration through the Combining of Feature Descriptors

445

in a Micro UAV SX2 (see Figure 2) (Senxis Innova-

tions in Drone Ltda, Uberl

andia, MG, Brazil) at an

average height of 100 meters. The datasets contains

respectively 565 (113 scenes and 5 channels), 670

(134 scenes and 5 channels) and 200 (40 scenes and 5

channels) images. Figure 3 shows an example image

scene containing all channels obtained by the Red-

Edge MicaSense camera coupled to the Micro UAV

SX2.

Figure 1: MicaSense Red-Edge camera by MicaSense.

Figure 2: Micro UAV SX2 by Sensix.

Figure 3: Example image scene containing all channels

(Blue, Green, Red, near-IR, red-edge respectively) obtained

by the Red-Edge MicaSense camera.

2.2 Image Registration

According to (Zitov and Flusser, 2003) image regis-

tration can be deﬁned as a process that overlaps two or

more images from various imaging devices or sensors

taken at different times and angles, to geometrically

align the images for analysis.

In this section, the authors present a review of the

main concepts related to the work developed in this

paper.

2.2.1 Feature Descriptors

Features, in the context of image registration, can be

deﬁned as a pattern that occurs in one location of the

image and differs from its closest neighbors. Usually

this pattern is associated with a sudden change in one

or more properties of an image (e.g., texture, color or

intensity). These features may or may not be loca-

ted at the same location of the change and are usually

small areas of the image, corners or points. Descrip-

tors are obtained by performing some type of proces-

sing on the region where a feature is present (Kumar,

2014) (Tuytelaars and Mikolajczyk, 2008).

Several techniques for obtaining features descrip-

tors were proposed. In this work, as previously

described, the following techniques were analyzed:

Min Eigen Features (MEF) (Shi and Tomasi, 1994),

Kaze Features (Alcantarilla et al., 2012), Binary

Robust Invariant Scalable Keypoints (BRISK) (Leu-

tenegger et al., 2011), along with the combination of

the BRISK and MEF, KAZE and MEF, BRISK and

KAZE techniques. In the following, a brief descrip-

tion of these techniques is provided.

Min Eigen Features are obtained by using the Shi-

Tomasi Corner Detector algorithm. This algorithm

was proposed by (Shi and Tomasi, 1994) and is based

on the Harris Corner Detector (Harris and Stephens,

1988) algorithm with a small change in the selection

criterion. The change in the selection criterion occur-

red on the fact that while in the Harris Corner Detec-

tor, the eigenvalues are passed to a function that re-

turns the score for the determination as to whether the

analyzed pixel is or not a corner, in the algorithm of

Shi-Tomasi that function has been removed and only

the eigenvalues are considered. As discussed in (Shi

and Tomasi, 1994) this change proved to be experi-

mentally superior to the selection criterion proposed

in the Harris Corner Detector algorithm. In addition

to the best selection method, this algorithm is also in-

variant to illumination, scale and rotation changes.

The Kaze algorithm was proposed by (Alcantarilla

et al., 2012) for the purpose of detecting 2D featu-

res in a nonlinear scale space to obtain greater accu-

racy of localization and distinctiveness. The Gaussian

blurring method used to generate the space scale in

other algorithms does not maintain the natural edges

of the analyzed image and also the noise is smoothed

at all scaling levels. In order to solve this problem,

the KAZE algorithm uses non-linear diffusion ﬁlte-

VISAPP 2019 - 14th International Conference on Computer Vision Theory and Applications

446

ring in conjunction with the Additive Operator Split-

ting (AOS) (Andersson and Marquez, 2016) method.

This algorithm is also invariant to illumination, scale

and rotation changes.

The Binary Robust Invariant Scalable Keypoints

(BRISK) was proposed by (Leutenegger et al., 2011)

in order to be an algorithm with a high performance,

however, with a drastic reduction of computatio-

nal cost when compared to algorithms like SIFT or

SURF. To obtain the localization of characteristics,

the BRISK algorithm uses the AGAST Corner Detec-

tor (Mair et al., 2010) technique, which holds a per-

formance improvement over the FAST algorithm. To

deal with scale changes, the BRISK algorithm ﬁnds

the points of interest within a space of scales, ap-

plying the technique of non-maximum suppression

(NMS) and by performing an interpolation between

all the scales. The BRISK algorithm is invariant to

scale and rotation changes.

The combination between techniques is accom-

plished by extracting all the features of both ima-

ges with the two techniques to be combined. Sub-

sequently, the feature vectors are vertically concate-

nated forming only one vector. Following this, the

removal of duplicate features is performed.

2.3 Registration Framework

The authors here used the data-driven framework for

multispectral image registration proposed by (Yasir,

2018), and summarized in Figure 4. Two chan-

ges were made in step 2 of this framework. The

ﬁrst change is due to the fact that this work aims

at the analysis of the image registry using only one

technique (be it single or a combination of techni-

ques). Therefore, after the extraction of the control

points, a graph will be constructed for each techni-

que described above. The second is the removal of

equal control points and the removal of outlier points

via random sample consensus (RANSAC) algorithm

(Fischler and Bolles, 1981).

This framework consists of the construction of a

complete graph where the nodes of the graph are the

channels to be registered and the weights the quan-

tity control points obtained by the algorithms between

those channels (see Figure 5). Then, using the Krus-

kal (Kruskal, 1956) algorithm, a Maximum Spanning

Tree is constructed. To ﬁnd the channel to be used

as the target for the other alignments, the weights

between the nodes are replaced by 1 and the Floyd-

Warshall all-pairs-shortest-path (Floyd, 1962) algo-

rithm is used. The node with the smallest sum of dis-

tances from itself to all the other nodes is selected as

the target channel for the registration scheme.

Figure 4: Registration Scheme proposed by (Yasir, 2018).

Figure 5: Example of a complete graph generated by the

framework.

With the aim to perform the comparison between

techniques, the authors use the number of distinct

control points found after the removal of the outlier.

The number of control points to be detected by each

technique are not limited.

An example of a result of the framework described

above is demonstrated in ﬁgure 6.

Detection of Control Points for UAV-Multispectral Sensed Data Registration through the Combining of Feature Descriptors

447

Figure 6: Example of output from Registration Scheme.

3 EXPERIMENTS

Each technique, or combining, (MEF, BRISK, KAZE,

MEF and BRISK, MEF and KAZE, BRISK and

KAZE) was applied individually over the datasets

and the results obtained are presented in this section.

As previously described, in this work the techniques

are being evaluated by the number of control points

found.

3.1 Soybean Dataset 01

Soybean Dataset 01 consists of 113 scenes in ﬁve dis-

tinct channels (blue, green, red, near-IR and red-edge

) resulting in a total of 565 images. Some images of

this dataset have few elements that favor the process

of alignment between the bands (e.g. plantations li-

nes and roads), it can be seen in the ﬁgure 7. For

this dataset the technique that generated the highest

average number of control points was Kaze Features

(see ﬁgure 8). This technique recognized on average

459 points on each image. The schema for the regi-

stry after all steps of the framework is shown in Fi-

gure 9. As previously described, this framework ge-

nerates a scheme for registration between the chan-

nels of an image, in order that the greatest number

of control points between the channels are obtained.

This scheme also demonstrates that the red and near-

IR channels have to be ﬁrst registered in the blue and

red-edge channels respectively, and then later recor-

ded in the green channel.

3.2 Soybean Dataset 02

Soybean Dataset 02 consists of 134 scenes in ﬁve dis-

tinct channels (blue, green, red, near-IR, red-edge)

resulting in a total of 670 images. Differently from

dataset 01, the images present in this dataset present

Figure 7: Example image of Soybean Dataset 01.

Figure 8: Average points per image in dataset 01.

Figure 9: Dataset 01 registration scheme using the KAZE

technique.

several elements that favor the alignment process be-

tween the bands (eg planting lines, trees, roads, plan-

ting failures), an image exemplifying these elements

can be seen in the ﬁgure 10. Due to these characteris-

tics, a higher number of control points was obtained

when compared to the other datasets. For this dataset

the best from among the techniques, when evaluating

the average number of control points, was obtained

VISAPP 2019 - 14th International Conference on Computer Vision Theory and Applications

448

throught the combining of those features obtained by

the KAZE technique with those from the MEF techni-

que (see ﬁgure 11). A total of 2406 control points

were recognized for each image. The large number

of control points available in the dataset images allo-

wed algorithms such as MEF to identify features not

detected by the KAZE technique. For this reason, the

combination of features obtained by MEF and KAZE

achieved a superior result. The schema for the regis-

tration after all framework steps is shown in Figure

12. This scheme demonstrates, as in Dataset 01, that

red and near-IR channels have to be ﬁrst registered in

the blue and red-edge channels respectively and sub-

sequently registered in the green channel.

Figure 10: Example image of Soybean Dataset 02.

Figure 11: Average points per image in dataset 02.

3.3 Soybean Dataset 03

Soybean Dataset 03 consists of 40 scenes in ﬁve dis-

tinct channels (blue, green, red, near-IR, red-edge) re-

sulting in a total of 200 images. Dataset 03 presents

only a few planting lines to be used by the algorithms

as control points. An image exemplifying the data-

set 03 is shown in the ﬁgure 13. For this dataset we

have the technique that generated the highest average

Figure 12: Dataset 02 registration scheme using the

MEF+KAZE technique.

Figure 13: Example image of Soybean Dataset 03.

number of control points as being Kaze Features (see

ﬁgure 14). This technique recognized on average 400

points on each image. The schema for the registry

after the conclusion of all steps in the framework is

shown in Figure 15. This structure presents a change

in the resulting scheme for datasets 01 and 02, unlike

the previous schemes, the near-IR channel is no lon-

ger be registered in the red-edge channel. All chan-

nels, with the exception of red, which continues to be

registered in blue, is recorded directly on the target

channel. For this dataset, just as with the prior sets,

the target channel is green.

4 CONCLUSION

In this work was explored the application of different

techniques for the registration of multispectral ima-

ges obtained by UAVs in soybean plantations. The

authors also evaluated how the combination of featu-

res obtained by different descriptors impacts on the

number of control points obtained.

As seen in the experiments section, on average,

the KAZE technique provided more control points for

Detection of Control Points for UAV-Multispectral Sensed Data Registration through the Combining of Feature Descriptors

449

Figure 14: Average points per image in dataset 03.

Figure 15: Dataset 03 registration scheme using the KAZE

technique.

the registration process than other techniques. In Da-

taset 02, the combination of the features obtained by

the MEF and KAZE techniques achieved the greatest

number of control points among the others techni-

ques. However, throught an analysis of the difference

obtained between the results, when compared to ot-

her factors (e.g. complexity to obtain features of two

different techniques), one concludes that the KAZE

technique was superior to other techniques evaluated

in this paper.

One of the factors that led to the KAZE techni-

que to obtain a large number of control points than

MEF and BRISK techniques is mainly due to how the

KAZE technique performs the scale invariance pro-

cess. The KAZE algorithm detects features in a non-

linear scale space by means of nonlinear diffusion ﬁl-

tering. Thus, the KAZE algorithm makes blurring lo-

cally adaptive to the image data, reducing noise but

retaining object boundaries, while obtaining superior

localization accuracy and distinctiveness (Alcantarilla

et al., 2012). This is quite different to BRISK that

uses Gaussian blurring, which, according to (Alcan-

tarilla et al., 2012), does not respect the natural boun-

daries of objects and smooths to the same degree both

the details and noise, reducing localization accuracy

and distinctiveness. The MEF algorithm, although

invariant to scale, presents a considerable degrada-

tion in the repeatability of the features when the scale

changes, as demonstrated by (Schmid et al., 2000).

The scale invariance is extremely important when we

are evaluating images obtained by UAVs since, as pre-

viously demonstrated, there are several factors (e.g.

wind speed, wind direction) that inﬂuence the path of

the UAV leading to distortions of scale and rotation.

Noteworthy is that althought dataset 2 presents a

small modiﬁcation, the scheme for channel registra-

tion is very similar among all datasets. The green

channel is always used as the ﬁnal target for the other

channels, where red and near-IR are ﬁrst recorded in

the blue and red-edge channels respectively. Through

a comparison of the registration scheme proposed by

(Banerjee et al., 2018), the spectral order (blue, green,

red, red-edge, NIR) was veriﬁed as not being the best

in any of the datasets evaluated.

The main limitation of this work is to have evalu-

ated only the number of control points obtained be-

tween channels. The quality of the alignment was

not evaluated in this work due to the lack of previ-

ously aligned datasets containing multispectral ima-

ges obtained by UAVs at low/medium altitude (100 to

400m). Satellite images cannot be used for this pur-

pose due to particular characteristics, such as higher

spatial resolution and greater stability. The datasets

presented in this work are manually aligned for ana-

lysis purposes.

Future work proposals are based on the evaluation

of bioinspired algorithms (e.g. Genetic Algorithms,

Swarm Intelligence, etc.) to perform the alignment of

multispectral crop images. Also, as previously descri-

bed, another study proposal that is under development

is the creation of an aligned dataset containing mul-

tispectral images obtained by UAVs to evaluate the

algorithms.

ACKNOWLEDGEMENTS

The authors gratefully acknowledges CAPES (Coor-

dination for the Improvement of Higher Education

Personnel), CNPq (National Council for Scienti-

ﬁc and Technological Development, Brazil) (Grant

#302416/2015-3), FAPEMIG (Foundation to the Sup-

port of Research in Minas Gerais (Grant #APQ-

03437-15) for the ﬁnancial support and the company

Sensix (http://sensix.com.br) for providing the images

used in the tests.

VISAPP 2019 - 14th International Conference on Computer Vision Theory and Applications

450

REFERENCES

Alcantarilla, P. F., Bartoli, A., and Davison, A. J. (2012).

Kaze features. In Fitzgibbon, A., Lazebnik, S., Pe-

rona, P., Sato, Y., and Schmid, C., editors, Computer

Vision – ECCV 2012, pages 214–227, Berlin, Heidel-

berg. Springer Berlin Heidelberg.

Andersson, O. and Marquez, S. R. (2016). A comparison

of object detection algorithms using unmanipulated

testing images: Comparing sift, kaze, akaze and orb.

Master’s thesis.

AUVSI (2013). The economic impact of unmanned aircraft

systems integration in the united states march. Techni-

cal report, The Association for Unmanned Vehicle Sy-

stems International.

Banerjee, B. P., Raval, S. A., and Cullen, P. J. (2018). Alig-

nment of uav-hyperspectral bands using keypoint des-

criptors in a spectrally complex environment. Remote

Sensing Letters, 9(6):524–533.

Berni, J. A. J., Zarco-Tejada, P. J., Suarez, L., and Fere-

res, E. (2009). Thermal and narrowband multispectral

remote sensing for vegetation monitoring from an un-

manned aerial vehicle. IEEE Transactions on Geos-

cience and Remote Sensing, 47(3):722–738.

Faria, L. C., R. L. F. M. J. F. (2018). Cell classiﬁcation using

handcrafted features and bag of visual words. In Pro-

ceedings of XIV Workshop de Vis

ao Computacional.

Fischler, M. A. and Bolles, R. C. (1981). Random sample

consensus: A paradigm for model ﬁtting with appli-

cations to image analysis and automated cartography.

Commun. ACM, 24(6):381–395.

Floyd, R. W. (1962). Algorithm 97: Shortest path. Com-

mun. ACM, 5(6):345–.

Harris, C. and Stephens, M. (1988). A combined corner

and edge detector. In In Proc. of Fourth Alvey Vision

Conference, pages 147–151.

Hunt, E. R., Hively, W. D., Fujikawa, S. J., Linden, D. S.,

Daughtry, C. S. T., and McCarty, G. W. (2010). Acqui-

sition of nir-green-blue digital photographs from un-

manned aircraft for crop monitoring. Remote Sensing,

2:290–305.

Junior, J. D. D., Backes, A. R., Abdala, D. D., and Es-

carpinati, M. C. (2018). Assessing the adequability

of image feature descriptors on registration of uav-

multispectral sensed data. In Proceedings of XIV

Workshop de Vis

ao Computacional.

Kruskal, J. (1956). On the shortest spanning subtree of a

graph and the traveling salesman problem. In Pro-

ceedings of the American Mathematical Society, pa-

ges 48–50.

Kumar, R. S. M. (2014). A survey on image feature des-

criptors.

Leutenegger, S., Chli, M., and Siegwart, R. Y. (2011).

Brisk: Binary robust invariant scalable keypoints. In

2011 International Conference on Computer Vision,

pages 2548–2555.

Mair, E., Hager, G. D., Burschka, D., Suppa, M., and Hi-

rzinger, G. (2010). Adaptive and generic corner de-

tection based on the accelerated segment test. In Da-

niilidis, K., Maragos, P., and Paragios, N., editors,

Computer Vision – ECCV 2010, pages 183–196, Ber-

lin, Heidelberg. Springer Berlin Heidelberg.

Sankaran, S., Khot, L. R., Espinoza, C. Z., Jarolmasjed, S.,

Sathuvalli, V. R., Vandemark, G. J., Miklas, P. N., Car-

ter, A. H., Pumphrey, M. O., Knowles, N. R., and

Pavek, M. J. (2015). Low-altitude, high-resolution

aerial imaging systems for row and ﬁeld crop phe-

notyping: A review. 70:112–123. Exported from

https://app.dimensions.ai on 2018/11/20.

Schmid, C., Mohr, R., and Bauckhage, C. (2000). Evalua-

tion of interest point detectors. International Journal

of Computer Vision, 37(2):151–172.

Shi, J. and Tomasi (1994). Good features to track. In 1994

Proceedings of IEEE Conference on Computer Vision

and Pattern Recognition, pages 593–600.

Soares, G., Abdala, D., and Escarpinati, M. (2018). Planta-

tion rows identiﬁcation by means of image tiling and

hough transform. pages 453–459.

Tuytelaars, T. and Mikolajczyk, K. (2008). Local invariant

feature detectors: A survey. Found. Trends. Comput.

Graph. Vis., 3(3):177–280.

Yasir, R. (2018). Data driven multispectral image registra-

tion framework. Master’s thesis, University of Saska-

tchewan.

Zecha, C. W., Link, J., and Claupein, W. (2013). Mobile

sensor platforms: categorisation and research appli-

cations in precision farming. Journal of Sensors and

Sensor Systems, 2(1):51–72.

Zitov, B. and Flusser, J. (2003). Image registration methods:

a survey. Image and Vision Computing, 21(11):977 –

1000.

Detection of Control Points for UAV-Multispectral Sensed Data Registration through the Combining of Feature Descriptors

451