Semi-automatic Analysis of Huge Digital Nautical Charts

of Coastal Aerial Images

Matthias Vahl

, Uwe von Lukas

, Bodo Urban

and Arjan Kuijper

Fraunhofer IGD, Joachim-Jungius-Str. 11, Rostock, Germany

Fraunhofer IGD, Fraunhoferstr. 5, Darmstadt, Germany

Keywords:

Semi-automatic Analysis, Electronic Nautical Charts, Coastal Aerial Images.

Abstract:

Geo-referenced aerial images are available in very high resolution. The automated production and updating

of electronic nautical charts (ENC), as well as other products (e.g. thematic maps), from aerial images is a

current challenge for hydrographic organizations. Often standard vision algorithms are not reliable enough

for robust object detection in natural images. We thus propose a procedure that combines processing steps

on three levels, from pixel (low-level) via segments (mid-level) to semantic information (high level). We

combine simple linear iterative clustering (SLIC) as an efﬁcient low-level algorithm with a classiﬁcation based

on texture features by supported vector machine (SVM) and a generalized Hough transformation (GHT) for

detecting shapes on mid-level. Finally, we show how semantic information can be used to improve results from

the earlier processing steps in the high-level step. As standard vision methods are typically much too slow

for such huge-sized images and additionally geographical references must be maintained over the complete

procedure, we present a solution to overcome these problems.

1 INTRODUCTION

Nowadays, the production of digital nautical charts

and paper charts is partially carried out from a com-

mon data base. By this, redundancies are avoided and

the production of thematic charts is accelerated. How-

ever, the ﬁrst acquisition as well as the update of this

data base from geo-referenced aerial images are both

still time- and cost-consuming manual processes. In

this paper we show how one can use computer vision

tools for a semi-automatic analysis of aerial images.

The basic idea is to develop a processing procedure

consisting of ﬁve processing steps. Simple image pro-

cessing techniques which detect single object are un-

suitable for this complex input data. It is necessary to

use domain-speciﬁc knowledge by model-based ap-

proaches. Especially combining low-level vision for

detection and high-level approaches for interpretation

seems to be promising. In this paper we present the

following contributions to tackle these problems:

1. We propose an efﬁcient and effective segmenta-

tion and recognition method for areas and areal

objects (areas like land and water) and particularly

shaped objects (parametric objects like groins or

piers).

2. We use information from inter-object relation-

ships to improve the recognition rate

3. We deal with process-speciﬁc problems, namely

data import and preparation (import and prepara-

tion of huge data amounts so that computer vision

algorithms can access them efﬁciently) and data

export and geo-referencing (we are able to store

detected objects and can keep the geo-references

– or make a reverse transformation).

After a discussion of related work we present our

approach containing the aforementioned three levels.

It is followed by an evaluation and the conclusions.

2 RELATED WORK

Bicego et al. present a color separation method within

the HSI (Hue, Saturation, Intensity) color space com-

bined with a following region growing with hystere-

sis threshold mechanism (Bicego et al., 2003). By

using the HSI color space a good separation between

chominance (hue and saturation) and intensity can be

reached. These channels are more concise. After-

wards, they use a very conservative threshold to seg-

ment the hue channel. This initial segmentation is the

100

Vahl M., von Lukas U., Urban B. and Kuijper A..

Semi-automatic Analysis of Huge Digital Nautical Charts of Coastal Aerial Images.

DOI: 10.5220/0005301501000107

In Proceedings of the 10th International Conference on Computer Vision Theory and Applications (VISAPP-2015), pages 100-107

ISBN: 978-989-758-091-8

 2015 SCITEPRESS (Science and Technology Publications, Lda.)

input for a region growing algorithm. They achieve

acceptable results with a probabilistic contour track-

ing approach to detect roads in housing developments

areas. To detect roads and other contours (Isard and

Blake, 1998) propose a probabilistic tracking based

on a modiﬁed jet stream resp. condensation approach.

Furthermore, (Xiao et al., 2008) present an approach

for extracting and applying a semantic layer from

low-height aerial images. The extraction of roads

without a-priori-knowledge is a big challenge. Thus

they use additionally correlated GIS information. The

registration step is done by a histogram correlation.

Whereas Xiao et al. try a model-based approach for

object extraction, (Straub et al., 2000) and (Ogawa

et al., 2000) use access to given maps and GIS in-

formation. (Letitia and Monie, 2008) show the seg-

mentation of satellite images by using adaptive neigh-

borhood mathematical morphology, which works on

multi-scale images. Moreover, (Galindo and Moreno,

2009) apply the algorithm of Otsu on a single color

channel to detect swimming pools from aerial images.

Afterward a reﬁnement of the result is done by using

active contours models (Kass et al., 1988). (Sirmacek

and Unsalan, 2008) describe how to use Otsu’s algo-

rithm to automatically detect red roofs. Furthermore,

they describe how shadows cause conspicuous pat-

terns in the blue channel. Before they use the thresh-

old, an ampliﬁcation is applied. (Vahl and von Lukas,

2013) adapt the blackboard metaphor introduced in

(Velthuijsen, 1992), for an semi-automated recogni-

tion of fac¸ades from oblique aerial images. Various

detection and processing algorithms – the experts –

are semantically coupled by the domain-speciﬁc grid

graph data model – the blackboard.

The algorithms can logically be combined but

have only weak dependencies between each other. In

summery, none of these approaches is able to solve

the problems mentioned in the Introduction com-

pletely. We thus present a procedure to couple meth-

ods on low to high level computer vision tasks to

analyze these huge digital nautical charts based on

coastal aerial images.

3 OUR APPROACH

To solve the sketched problems we propose a process-

ing procedure (see Figure 1) with the following ﬁve

steps:

1. In the pre-processing step image data is prepared

to be suitable for next 3 computer vision steps.

2. The low-level step collects pixels into segments

3. The mid-level step merges segments into semantic

Figure 1: Our proposed processing procedure, consisting

of a pre-processing step, three computer vision steps, and a

post-processing step.

objects

4. The high-level step interprets inter-object rela-

tionship to improve earlier steps

5. In the post-processing step results from the vision

processing steps 2–4 are transformed back to geo-

graphical coordinates and recognized objects will

be serialized into a geographical vector format.

Starting point of the whole process are aerial

images. Throughout this paper we use those pro-

vided by Bundesamt f

ur Geod

asie und Karthographie

(BSH, German Federal Agency for Cartography and

Geodesy). They have a resolution of 20 cm per pixel

(px) as ECW (Enhanced Compressed Wavelet) com-

pressed GeoTIFF. The coast of Germany is divided by

the BSH into nautical charts. The sample GeoTIFF

we use in the examples is the chart ”Elbm

undung”,

where the Elbe river ﬂows into the North Sea. It has

a size of 271 977×187 340 px

which is equivalent

to 54×37 km

, or approx. 51 Gpx. Because of this

huge size, standard computer vision algorithms will

not work properly on this data. A reorganization of

the data is thus necessary.

The second step of the process chain is the low-

level step that works on pixel level. The entropy of

one pixel is very low in opposite to their quantity. So a

fast algorithm for collecting pixels with similar prop-

erties is needed. Results of this step are segments.

Important is that one segment belongs completely to

one object. Oversegmentation – one object consists

of several segments – is unproblematic.

Segments build the starting point for mid-level al-

gorithms. Their task is to combine segments to ob-

jects. An object is the smallest exportable informa-

tion unit. Objects have a big morphological variance.

There are objects that represent an area such as water

or grassland by multi-polylines, but also by more or

less simple shapes, such as rectangles, circles.

The most complex step of the process chain –the

high-level step– aims at bringing detected objects into

relation and resolving ambiguities by including con-

Semi-automaticAnalysisofHugeDigitalNauticalChartsofCoastalAerialImages

101

text information. In this step groups of objects (e. g.

groins or wind farms) can be recognized by using

characteristic properties (distance to each other, di-

rection, shape).

The last process step –the post processing step–

transforms the positions of known objects from the

image space coordinate system. This step is necessary

because of the image processing algorithms working

on the smaller, Cartesian image space. The target

systems need geographical referenced objects, e. g.

WGS 84. Finally the computer-internal representa-

tion of detected objects is transformed into a neutral

exchange format like the IHO S-57, a standard of In-

ternational Hydrographic Organization (IHO) for the

digital exchange of nautical, hydrographic and bathy-

metric information.

3.1 Pre Processing

Taking into account that aerial images are very data

intensive it is necessary to divide them into smaller

pieces. Here, we use image pyramids with different

resolution layers consisting of 1024×1024 px

-sized

tiles, see Figure 2.

Figure 2: Image pyramid. Resolution decreases from bot-

tom to the top.

In addition, we later on need to merge partial ob-

jects of bigger objects that lay in two or more tiles.

This size enables a (MATLAB) script to work for pro-

totyping with acceptable response times. While di-

viding large data sets into pieces it is important to

keep geo-references. This is done by using some tools

from the Geospatial Data Abstraction Library Project

(GDAL).

Table 1 shows resolution, edge length of tiles, as

well as average memory usage. At the highest pyra-

mid level the data volume is shrunken by a factor of

16, memory usage as calculated by a factor of 256 but

in real by (“only”) a factor 232. The computation time

saved depends on the complexity of the algorithms.

3.2 The Computer Vision Steps

As shown in Figure 1 and listed above, the proposed

computer vision processing steps can be divided into

three steps: collecting pixels to segments, merging

segments into semantic objects and interpreting inter-

object relationship to improve earlier steps. In the fol-

lowing sections we describe each step in detail.

3.2.1 The Low-level Step (Pixels to Segments)

As already mentioned in the related work section,

threshold-based methods work on the distribution (in-

tensity of gray level or one or more color channel)

of pixels. Methods that work more globally were ne-

glected because of the different exposure of the source

material, the low contrast, and the large number of ob-

ject classes. Hue-based methods for water detection,

proposed in the literature, e.g. (Galindo and Moreno,

2009), do not work for the North Sea conditions. For

shallow water the blue part is too low, for dried-up

mud ﬂats it is absent. But also deep water of e.g.

the harbors has not enough intensity in the blue chan-

nel, because of the enrichment with sediments. It thus

can not be detected by classical water detection meth-

ods. Therefore, hue-based approaches using a global

threshold are not suitable. Furthermore, objects from

different object classes differ not only in the color

property but also in color and texture.

For these reasons we argue that it is more pro-

ductive to apply a region-based segmentation which

is working with local thresholds. Typical representa-

tives of this algorithm family are the watershed ap-

proach (Huguet et al., 2004) or – a little bit newer –

the simple linear iterative clustering (SLIC) approach

(Achanta et al., 2010). In literature (e.g. (Achanta

et al., 2012)) SLIC is rated better than other ap-

proaches. It is available as C++ library (Vedaldi and

Fulkerson, 2008) for different platforms and many

programming language bindings, especially MAT-

LAB. Ozden et al. (Ozden and Polat, 2007) show a

promising approach using color, texture and space in-

formation for segmentation, but unfortunately no im-

plementation is provided. With a given region size

and regularization behaviour the contours of the rel-

evant object are exactly traced. The resulting over-

segmentation is not problematic and will be treated

in the next step – the mid-level step. SLIC uses hue

information only implicitly. A results of SLIC seg-

mentation can be seen in Figure 3.

3.2.2 The Mid-level Step (Segments to Semantic

Objects)

The mid-level step depends on the kind of objects we

have to recognize. One can choose between areal ob-

jects, such as water or farmland, and shaped objects,

such as buoys, groins or other coastline constructions.

For the ﬁrst one the texture in the foreground is impor-

tant, whereas for the second one the shape is relevant.

VISAPP2015-InternationalConferenceonComputerVisionTheoryandApplications

102

Table 1: Resolution steps of the quadratic tiles. The average memory usage per kilometer for the given test area and using the

lossless compression LZW.

Level Resolution [cm/px] Edge length of the tile [m] Average memory usage per area [MB/km

]

0 20 205 58,00

1 40 410 16,00

2 80 820 4,00

3 160 1640 1,00

4 320 3300 0,25

Figure 3: The SLIC (Vedaldi and Fulkerson, 2008) algo-

rithm clusters pixels on base of their color and neighbour-

hood.

Areas – Areal Objects. The starting point for the

detection of areas are groups of pixels clustered into

segments (done in the previous low-level step). On

the one hand for every segment known texture fea-

tures exist and on the other hand important properties

are integrated by the clustering, e.g. the color. The as-

signment of segments to an object class and the merg-

ing of segments from the same class is done within

the classiﬁcation step. First, signiﬁcant features must

be selected. Figure 4 compares normalized Haralick

features (Haralick et al., 1973) of the land class to

features of water class. One can see that there are

signiﬁcant features, but also features with the same

characteristics. For a feature selection further statis-

tical properties must be taken into account, e. g. low

variance. In case of many classes an automated fea-

ture selection can be used, but in here it sufﬁces to

choose the features manually. Based on Figure 4 a

good choice is to take feature 1, 2, 3, 4, 5 and 9. For

classiﬁcation a supported vector machine (SVM) in

one-against-one mode is used. Due to the the low

number of classes the use of a decision tree is not nec-

essary.

Shaped Objects. In contrast to detecting areas, for

which texture features play an important role, for de-

tecting shape objects we additionally use contour in-

formation. We do not extract this from edge-detected

images because standard methods such as Canny get

problems with vast-structured images. We there-

fore build contours by merging neighboring segments

which belongs to the same class. Afterward we take

the contour of this objects for the further analysis.

Figure 4: Different characteristic of the normalized Hara-

lick features in four directions.

It is well-known that the Hough transformation is

limited suited because it requires a shape parametriza-

tion. Therefore, we use an own implementation of

the Generalized Hough Transformation (GHT) which

has the advance to work efﬁciently for parametrized

shapes. On the basis of the template similar objects

will be detected by GHT.

Tests with this approach are done for different

kinds of groins (see Figure 5). The results for standard

groins in resolution step 1 (40 cm/px, see Table 1) are

very satisfying. For buoys and other small objects we

have to use resolution step 0 (20 cm /px).

Reduction of the Search Space. Since the Hough-

Transformation is a kind of brute force approach, the

memory usage and the computing time is relatively

vast. Here speeding up by hardware (e.g. general pur-

pose graphical processing units), strategies that limits

the search space and optimization of the algorithms

can help.

The reduction of the space that has to process cal-

culation can accelerated massively, depending on the

object class. For example, buoys are only located on

water areas. It is further necessary to bring the ob-

ject classes in a hierarchical structure to enable fast

methods. Without knowledge about the coastal zone

however, this is impossible. We therefore introduce

the geometry type GT , which the S-57 object classes

classiﬁes by geometrical criteria (see Figure 6).

Search spaces S are deﬁned by space-deﬁning ob-

ject, such as water-land-line or meta data like harbour

region. For every geometry type GT , a probability

of occurrence p(GT,ST ) per search space type ST is

given. In the case that p(GT, ST ) > 0, for every space

Semi-automaticAnalysisofHugeDigitalNauticalChartsofCoastalAerialImages

103

Figure 5: Generalized Hough Transformation (Search

space: Scale s: 0.5 ≤ s ≤ 2 with increment 0.5 and rotations

ϕ: 0 ≤ ϕ ≤ 2π with increment π/10). Top: Hough space and

maxima. Middle: Segmented image with projected max-

ima. Bottom: Original image with projected maxima.

of the space type ST the search algorithm for geom-

etry type OT

,..., OT

is applied. This reduces the

calculating costs damatically, but it needs formalized

knowledge about the domain. Figure 6 shows the sort-

ing of object classes by shape.

Table 2 exemplary shows when the probability of

occurrence is null and which object in which region

can expected. The probability must be determined by

Figure 6: Sorting of S-57 object classes (selection) by shape

(geometry type) and context.

Figure 7: Graphical representation of semantic information.

There are relationships between buoys and theirs environ-

ment.

test data or must be estimated.

3.2.3 High-level Step (Objects to Scene)

One object class family that needs to be recog-

nized are buoys, especially lateral buoys (S-57 object

identiﬁer BOYLAT). These objects are very small.

Their resolution is of only a few pixels. They are

surrounded by water and usually have a signiﬁcant

shadow that also is surrounded by water. The rea-

son for this is that normally, aerial images are taken

on cloudless sunny days. The buoy shape maybe dis-

torted perspectively, depending on the distance to the

optical center. The size and direction of the shadow

depend on the illumination situation, especially the

weather and the solar altitude.

Buoys and shadow of buoys induce – in contrast

to the surrounding water – conspicuousness within the

color space. Corresponding to its color (green or red)

there is a signiﬁcant shift in the color channel resp.

within the saturation of the buoy’s shadow. This prop-

erty and relationship can visualized and problemati-

cally formulated (see Figure 7).

Figure 8 and Figure 9 show the results of the ex-

periments for two different S-57 object classes. The

detection works ﬁne and regarding the context the al-

gorithm can infer the object class. As result position

VISAPP2015-InternationalConferenceonComputerVisionTheoryandApplications

104

Table 2: Search space types.

1.6cmGeometry type GT Search space type ST

Coastal strip Water Harbour Land

rectangular groin p = 0 buildings, pontoons, cars

circular (small) buoy p = 0

circular (big) p = 0 silo/tank, cooling tower

irregular p = 0 ships ships, hulks p = 0

Figure 8: Detection of lateral buoys. Top: Aerial image af-

ter step 3 pixels to segments. Bottom: After step 4 segments

to objects. Conspicuousness within the color spectrum: low

saturation (shadow of the buoy), enhanced green (buoy).

Figure 9: Detection of land-side silos. Top: Aerial image

after step 3 pixels to segments. Bottom: After step 4 seg-

ments to objects.

we can use the base point of the buoy body or the mid-

dle point of the parting line between the buoy and the

buoy’s shadow. We transform this position from im-

age coordinates to geographical coordinates and take

the color as object attribute (e. g. in IMO S-57 BOY-

LAT: BOYSHP=?, COLOUR=green). Assuming that

the error in the image space is 15 px the maximum

deviation is between 1 and 3 meters. Buoys have a

speciﬁc shape respectively silhouette that allows the

sailor to identify the kind of buoy against poor light-

ning condition. The shape of the buoy cannot be re-

constructed since the perspective is nearly orthogonal

so that we only detect that the buoy is circular. Since

buoys are optimized for horizontal perspectives this is

not a problem.

3.3 Post Processing

In the post processing step we have to transform the

object position back from the image coordinate sys-

tem to a geographical coordinate system. Further-

more, we have to export the own internal data model

to a vector-oriented geo data exchange format.

4 EXPERIMENTS

First, we implemented our approach as a prove of con-

cept in MATLAB. Table 3 shows the accuracy of the

SLIC segmentation followed by an SVM classiﬁca-

tion based on the selected Haralick features. In this

matrix the values show how well classes can be sep-

arated. For example, all water segments are distin-

guished from land segments. To divide between road

and farmland2 the algorithm reaches 87 %, but be-

tween road and farmland1 it yields 55 % only, as they

are rather identical.

The recognition rate for areal objects is generally

between 60 and 100 % and for the relevant areas be-

tween 83 and 100 %. Only the separation of roads and

land is 78 %. For the examined shaped objects the de-

tection rate is more than 80 %. The exactness is very

high. On the best resolutions (level 0) on contours al-

most pixel exactness (20 cm) is reached for the testing

region. On lower resolution the correct segmentation,

Semi-automaticAnalysisofHugeDigitalNauticalChartsofCoastalAerialImages

105

Table 3: SLIC and SVM: Selection of Haralick features 1, 2, 3, 4, 5, 7 with indirectly regarding the hue value by the SLIC

algorithm. The important textures can be distinguished well to very well.

% Land Other Road Grassland Farmland1 Farmland2

Water 100 100 99 60 89 65

Land – 96 78 83 50 85

Other – 77 99 80 97

Road – 86 55 87

Grassland – 87 74

Farmland1 – 86

especially of small objects, fails frequently, because

some segments are too small for a texture analysis by

Haralick’s method.

Table 4: Number of segments in the disjunctive training

data set M

train

and test data set M

test

Class Number of Segments

train

| |M

test

| Σ

Water 22 133 155

Land 9 46 55

Other 14 3 17

Road 45 9 54

Grassland 299 121 420

Farmland1 72 39 111

Farmland2 477 32 509

Table 4 shows the number of segments in the dis-

junct training and test data sets.

Our classiﬁcation yields good to very good re-

sults for the assignment to object classes and objects.

Using texture features in combination with an SVM

makes this approach robust. The complexity is quite

high, because for every resolution level a feature set

must be selected and an SVM must be trained. For

this scenario we concentrated on fewer classes. For an

increasing number of classes we could use concepts

like decision trees. Calculating costs for determining

the texture features is quite high but can be optimized

by better feature selection and speed-up by hardware.

Calculating of features can be easily done in parallel.

5 CONCLUSIONS

We presented a method combining processing steps

on three levels, from pixel (low-level) via segments

(mid-level) to semantic information (high level) to

analyze geo-referenced aerial images. As these im-

ages typically have very high resolution, it is a prob-

lem to automatically produce and update digital nau-

tical charts from aerial images. Currently, the ﬁrst

acquisition as well as the update of data bases from

geo-referenced aerial images are both time- and cost-

consuming manual processes.

By combining SLIC as an efﬁcient low-level al-

gorithm via an SVM classiﬁcation based on texture

features and a GHT for detecting shapes on mid-level

with semantic information in the high-level step, we

obtained a reliable robust object detection that cur-

rent standard vision algorithms cannot deliver. We

propose an efﬁcient and effective segmentation and

recognition method for areas and areal objects and

particularly shaped objects. Furthermore we use in-

formation from inter-object relationships to improve

the recognition rate. Finally, we deal with process-

speciﬁc problems, namely data import and prepara-

tion and data export and geo-referencing.

Regarding segmentation and recognition of

coastal zone objects there are two important results:

(1) Choosing the appropriated pixel-based algorithms

with classiﬁcation by texture features is suitable for a

sharp separation of different areal objects and areas.

Our proposed detection of shaped objects like groins,

buoys, and piers is based on areal segmentation,

but additionally uses contours yielding good results.

Identiﬁcation is done by adapted GHT. Additionally,

shape features can be used to improve segmentation.

(2) The whole coastal zone can be seen as scene.

Context information from the scene can be used to

improve the fault tolerance of the object recognition

in the overall system. Furthermore, it is the key for

identiﬁcation of complex objects. For an automatic

recognition of images it is necessary to include the

context and to support the detection by a knowledge

base. Combination of image pyramids and geo-

speciﬁc tools allows for using computer vision on

huge data and for keeping geographical referencing

for an exact back projection of objects positions.

For future development it is planned to derive

domain-speciﬁc knowledge from semantic models for

the third vision step. In addition, we will try to use

probabilistic models on this high-level step to detect

composite objects and virtual objects such as water-

land-line or dykes.

MATLAB is very practical for rapid prototyping

and interactive testing, but the primary ”data base” is

the ﬁle system with images and thousands text ﬁles

of vectors and matrices. For this reason, and because

VISAPP2015-InternationalConferenceonComputerVisionTheoryandApplications

106

of the performance, we started to re-implement it in

C++ using as many as possible free 3rd party libraries.

Currently, the data model, the access to a lightweight

data base and some basic processes are already imple-

mented. Futhermore we have to speed-up expensive

operations such as the generalized Hough transforma-

tion by hierachical approaches and by using parallel

processing on GPGPUs.

ACKNOWLEDGEMENTS

The authors would like to thank Steffen Grammann,

Jana Vetter und Manuela Sch

onrock from the divi-

sion Nautical Information Service at BSH for their

support, knowledge, and discussions regarding dig-

ital nautical charts of coastal aerial images and the

processing procedures.

REFERENCES

Achanta, R., Shaji, A., Smith, K., Lucchi, A., Fua, P., and

Susstrunk, S. (2012). Slic superpixels compared to

state-of-the-art superpixel methods. IEEE Trans. Pat-

tern Anal. Mach. Intell., 34(11):2274–2282.

Achanta, R., Shaji, A., Smith, K., Lucchi, A. A., Fua, P., and

usstrunk, S. (2010). SLIC Superpixels. Technical

Report June, EPFL, EPFL.

Bicego, M., Dalﬁni, S., and Murino, V. (2003). Extrac-

tion of geographical entities from aerial images. 2nd

GRSS/ISPRS Joint Workshop on Data Fusion and Re-

mote Sensing over Urban Areas.

Galindo, C. and Moreno, P. (2009). Swimming pools lo-

calization in colour high-resolution satellite images.

Geoscience and Remote Sensing Symposium, 2009

IEEE International, IGARSS 2009, pages 510–513.

Haralick, R. M., Shanmugam, K., and Dinstein, I. (1973).

Textural features for image classiﬁcation. IEEE

Transactions on Systems, Man and Cybernetics.

Huguet, A., de Andrade, M. C., Carceroni, R. L., and

Araujo, A. d. A. (2004). Color-based watershed seg-

mentation of low-altitude aerial images. In Proceed-

ing of the XVII Brazilian Symposium on Computer

Graphics and Image Processing SIBGRAPI’04.

Isard, M. and Blake, A. (1998). Condensation—conditional

density propagation for visual tracking. International

journal of computer vision, 29(1):5–28.

Kass, M., Witkin, A., and Terzopoulos, D. (1988). Snakes:

Active contour models. International journal of com-

puter vision, 331:321–331.

Letitia, S. and Monie, E. C. (2008). Road segmenta-

tion from satellite aerial images by means of adaptive

neighborhood mathematical morphology. 2008 Inter-

national Conference on Computer and Communica-

tion Engineering, pages 427–432.

Ogawa, Y., Iwamura, K., and Kakumoto, S. (2000). Ex-

tracting object information from aerial images: a map-

based approach. IEICE TRANSACTIONS on Informa-

tion and Systems, 83(7):1450–1457.

Ozden, M. and Polat, E. (2007). A color image segmenta-

tion approach for content-based image retrieval. Pat-

tern Recognition, 40(4):1318–1325.

Sirmacek, B. and Unsalan, C. (2008). Building detec-

tion from aerial images using invariant color features

and shadow information. International Symposium on

Computer and Information Sciences, 2008. ISCIS ’08.

Straub, B., Wiedemann, C., and Heipke, C. (2000). Towards

the automatic interpretation of images for GIS update.

International Archives of International Society for

Photogrammetry and Remote Sensing, XXXIII:525–

532.

Vahl, M. and von Lukas, U. (2013). Toward Automated

Fac¸ades Generation from Oblique Aerial Images. Eu-

rographics Workshop on Urban Data Modelling and

Visualisation, pages 5–8.

Vedaldi, A. and Fulkerson, B. (2008). VLFeat: An Open

and Portable Library of Computer Vision Algorithms.

Velthuijsen, H. (1992). The nature and applicability of the

blackboard architecture. PTT research.

Xiao, J., Cheng, H., Han, F., and Sawhney, H. (2008). Geo-

spatial aerial video processing for scene understand-

ing and object tracking. In Computer Vision and Pat-

tern Recognition, 2008. CVPR 2008. IEEE Confer-

ence on, pages 1–8. Ieee.

Semi-automaticAnalysisofHugeDigitalNauticalChartsofCoastalAerialImages

107