Visual Navigation with Street View Image Matching

Chih-Hung Hsu

and Huei-Yung Lin

Department of Electrical Engineering, National Chung Cheng University,

168 University Rd., Chiayi 621, Taiwan

Department of Electrical Engineering, Advanced Institute of Manufacturing with High-Tech Innovation,

National Chung Cheng University, 168 University Rd., Chiayi 621, Taiwan

Keywords:

Visual Navigation Assistance, Image Matching, Panoramic Image.

Abstract:

The vision based navigation approach is a key to success for the driving assistance technology. In this work,

we presents a visual navigation assistance system based on the geographic information of the vehicle and

image matching between the online and pre-established data. With the rough GPS coordinates, we utilize the

image retrieval algorithms to ﬁnd the most similar image in the panoramic image database. The searching

results are then compared with the input image for feature matching to ﬁnd the landmarks in the panoramic

image. By using the 360

◦

ﬁeld-of-view of the panoramic images, the camera’s heading can be calculated by

the matching results. Finally, the landmark information is identiﬁed by the markers on the Google map as

visual guidance and assistance.

1 INTRODUCTION

The navigation techniques based on visual informa-

tion are getting more and more popular in recent

years. When people are in an unknown or unfamiliar

environment, visual navigation plays a key role to the

movement guidance. In general, landmarks are the

most reliable visual cues for the perception of outdoor

scenes. Using landmarks as references is very obvi-

ous to understand where you are and where do you

want to go. In the traditional navigation systems, a

number of available landmarks are tagged on the map

and their images are provided for location identiﬁca-

tion. Since the landmarks may appear differently ob-

served from different viewing directions, scene recog-

nition becomes a nontrivial task if the images are cap-

tured perspectively from a single viewpoint.

Currently, it is still not possible to achieve the

fully autonomous navigation due to various techni-

cal difﬁculties. Thus, many researchers are devoted

themselves to the assisted technologies for naviga-

tion. The previous work on the development of as-

sisted navigation techniques is generally divided to

two categories for the indoor and outdoor applica-

tions. In the indoor case, Saab et al. use the radio

frequency identiﬁcation device (RFID) for localiza-

tion (Saab and Nakad, 2011). Several passive tags

are pre-installed in the indoor environment, and the

RFID device can retrieve the information from the

tags when a user gets close to those places. Mulloni

et al. utilize the image marker to encode the infor-

mation in the proposed system (Mulloni et al., 2009).

The users can use their cell phones to capture and de-

code the markers to determine the locations. Huang et

al. present a vision based self-localization technique

for mobile robot applications (Huang et al., 2012). A

camera is used to capture the images of pre-installed

2D barcode patterns on the ceiling for indoor local-

ization. The above approaches require structured or

controlled environments to derive the location infor-

mation and therefore are not suitable for the outdoor

applications.

For the outdoor navigation, a primary information

source can be obtained from the images of the scenes.

The images are capable of providing the visual assis-

tance close to the human perception while not alter-

ing the environment. There have been many tech-

niques developed to construct the object classiﬁers

(i.e., trees, cars, buildings, roads, pedestrian, etc.) and

used to process the input scene images for the deriva-

tion of semantic maps. The most likely scenes in the

image database are then identiﬁed according to the

semantic maps (Wang et al., 2015; Yao et al., 2014;

Siddiqui and Khatibi, 2014). Some researchers as-

sume that the users are able to capture the most repre-

sentative objects such as buildings and deﬁne them as

landmarks. The navigation system is designed to use

the input image to match and ﬁnd the related land-

Hsu, C-H. and Lin, H-Y.

Visual Navigation with Street View Image Matching.

DOI: 10.5220/0005674405830589

In Proceedings of the 11th Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2016) - Volume 3: VISAPP, pages 585-591

ISBN: 978-989-758-175-5

585

mark image and return the corresponding information

to the users. Two commonly adopted methods, in-

cluding image retrieval and landmark matching, are

brieﬂy reviewed as follows.

The image retrieval approach is based on the im-

age features like texture, line, shape, etc. Turcot et al.

use the bag-of-words to construct an image database,

and choose the most useful features for image re-

trieval (Turcot and Lowe, 2009). Philbin et al. adopt

the random tree to construct the relationship between

the image and the image features, and then use it for

image retrieval (Philbin et al., 2007). Hu et al. pro-

pose a gradient ﬁeld HOG (GF-HOG) technique. The

users sketch the shapes of objects and the system can

extract the GF-HOG features to perform the image

retrieval (Hu and Collomosse, 2013). In the recent

years, the GPS system is extensively used in many

applications. Some researches also utilize the GPS

data to increase the accuracy of image retrieval (Chen

et al., 2011; Altwaijry et al., 2014; Zamir and Shah,

2010; Liu et al., 2012).

In the landmark matching techniques, Naoyuki et

al. use SURF (Speeded Up Robust Features) (Bay

et al., 2008) to perform image matching between the

panoramic and perspective images (Uchiyama et al.,

2009). However, due to the 360-degree ﬁeld-of-view

(FOV) of the panoramic images, good results can-

not be achieved using the traditional image features

(Lin et al., 2013). Chen et al. crop the landmarks in

the panoramic image to many blocks, calibrate these

sub-image blocks, and then use these sub-images to

match with the input image (Chen et al., 2011). Sim-

ilarly, the traditional 2D orientation descriptor such

as SIFT (Scale Invariant Feature Transform) is also

unsuitable for panoramic image because of the image

distortion and camera rotation (Lowe, 2004). Ben-

jamin et al. propose a 3D orientation SIFT and use it

to improve the matching results for panoramic images

(Resch et al., 2014).

In the past few years, Google Map Street View

could be the most famous image guided navigation

system. The users only need to input theirs GPS co-

ordinates, and the system will quickly return the loca-

tions on the map and the surrounding 360-degree im-

ages. Since the GPS accuracy is occasionally affected

by the weather condition and other factors, the users

should manually check the images are correct or not.

In this work, we present a visual navigation assistance

system. The user only have to input a landmark image

and the nearby GPS coordinates, the proposed navi-

gation system is able to return the name of the identi-

ﬁed landmark. Due to the advances of smart wearable

devices, it becomes possible to obtain the GPS en-

abled images easily. Our navigation assistance tech-

Figure 1: An overview of the proposed visual navigation

assistance framework.

nique utilizes the image and geographical information

to search in the panoramic image database. Since the

panoramic image contains 360

◦

FOV, the matching

result can be used to calculate the heading of the im-

age capture orientation. The corresponding landmark

marker is then obtained by searching the Google map

images with the heading angle.

2 PROPOSED FRAMEWORK

The proposed framework consists of three stages: (i)

image retrieval from the panoramic image database,

(ii) image matching between the perspective and

panoramic images, (iii) calculation of the camera

heading and using the angle to search the landmark

marker on the Google Map. The overview of our sys-

tem is shown in Figure 1. We will describe the method

used in detail below.

2.1 Database and Image Retrieval

In the proposed navigation technique, several image

databases are ﬁrst recorded in different regions. Each

image database consists of panoramic images and the

associated geographical coordinates. For testing or

vehicle localization, we capture the perspective im-

ages using a conventional camera and also record the

geographical information around the regions. The test

images are acquired with the landmarks we are inter-

ested for navigation assistance.

For an input perspective image, our navigation

system should identify the most similar panoramic

image in an image database. Since our database con-

tains images and the associated geographical coordi-

nates, there are two simple ways to perform the image

retrieval– using either the image based techniques or

the GPS information. These two methods, however,

both have some drawbacks. First, using the image

VISAPP 2016 - International Conference on Computer Vision Theory and Applications

586

Figure 2: The nodes assigned by the images in the database.

based technique alone will require a lot of search-

ing time since there usually have many street scene

images in the database. Second, the GPS informa-

tion is not always reliable because it may be affected

by some factors, such as bad weather conditions, tall

building, tunnel, etc.

Our navigation assistance technique combines the

image based approach and geographical coordinates

to construct the image retrieval system. We capture

the panoramic images on the left and right lanes along

a road for every 13 meters. A node is assigned to 6

neighboring panoramic image capture locations as il-

lustrated in Figure 2. The nodes of 6 images plus 2 ad-

ditional ones in the front and 2 additional ones in the

back are used for image retrieval and there is no need

to record the GPS coordinates for each panoramic im-

age. Given an input image, its location will be associ-

ated with a nearest node which is veriﬁed by feature

matching on the panoramic images in the database.

In the previous research, it is shown that SIFT

outperforms other features (Juan and Gwun, 2009).

When the 10 candidate panoramic images are found,

the number of SIFT matching points is considered as

the similarity measure between the panoramic and in-

put images. We extract the SIFT feature points in the

input image and 10 candidate panoramic images. The

features of candidate images are merged to a feature

vector

F = { f

, f

pre1

, f

pre2

, f

back1

, f

back2

}

The feature vector F is used to match with the input

image feature f

. A score is given to each panoramic

image according to the number of matching feature

points. The panoramic image which has the highest

score is the most similar to input image. Mathemati-

cally, it can be written as

, . . . , S

, S

pre1

, S

pre2

, S

back1

, S

back2

} = M( f

, F)

= argmaxM( f

, F)

where I

is the retrieved panoramic image.

(a) (b)

Figure 3: The image segmentation to separate the sky and

landmark regions. (a) The original image. (b) The edge

image obtained by Canny edge detection. (c) The result

from dilation and erosion. (d) The segmentation result by

the FloodFill algorithm.

2.2 Landmark Matching

Although the most similar image is found in the im-

age retrieval stage, the system still does not know

the corresponding landmark in the panoramic im-

age. The SIFT features are commonly used for land-

mark matching, but the results we ﬁnd are generally

not good enough. There always exist some incor-

rect matching points. Since the landmarks of inter-

ests usually contain regular appearance, we add land-

mark contour line features to enhance the correctness

of SIFT feature matching. It is the common case that

the landmarks in the images are connected with the

sky. Thus, the landmark contour lines can be ob-

tained by separating the sky region from the fore-

ground. We use the segmentation method proposed

by Laungrungthip et al. (Laungrungthip et al., 2008).

Canny edge detection (Canny, 1986) is ﬁrst applied

to extract the line segments, followed by the dila-

tion and erosion operations to ﬁll the gaps in broken

edges. The FloodFill algorithm is then carried out to

ﬁll the region from the top of the image (Hughes et al.,

2013). Finally, the landmarks are separated from the

sky by the contour lines detected by Hough transform

(Hough, 1962).

In the outdoor scenes, the captured images usu-

ally contain a lot of trees and leaves. These objects

should be treated as noise because it will cause un-

reliable landmark matching. In general, the line seg-

ments associated with trees and leaves have smaller

length and the most representative line segments have

longer length. According to this property, the noise is

ﬁltered out by comparing the length of line segments.

We can then choose the longest line segments at dif-

ferent angles to be the most representative ones. The

generation of line segments is based on the fact that

Visual Navigation with Street View Image Matching

587

Figure 4: The ROI setting for the feature descriptor con-

struction.

there is a signiﬁcant color change on both sides of

the line. If we choose two ROIs from each side of a

line (as shown in Figure 4) and consider these ROIs

as a single region, its histogram will have two peak

values– associated with the sky and the landmark, re-

spectively.

We construct the feature descriptor of a line seg-

ment according to the properties, and denote it as a

13-dimensional feature vector

= [ L

, H

max1

, H

max2

, V

max1

, V

max2

, H

avg1

, H

avg2

avg1

, V

avg2

, D

v max

, D

v avg

, H

sd1

, H

sd2

]

where L

is the angle of of the line, H

max

and V

max

are

the peak values of the hue and intensity histograms

of the ROIs, H

avg

and V

avg

are the average values of

the hue and intensity histograms of the ROIs, D

v max

and D

v avg

are the difference of the ROI’s maximum

and average values, H

is the standard deviation. The

feature descriptor is used to match the line segments

between the input and panoramic images. We use L

to ﬁlter out the segments which have a large difference

of line angles, and calculate the Euclidean distance

using the rest 12 parameters.

There are usually many matched line segments

and the number should be further reduced. In this

stage, the segments with the top 3 similarity scores are

chosen as candidates. Any two line segments in the

input image will be merged if the distance between

their midpoints is smaller than a threshold. The mid-

point of the merged line segment is then considered as

a possible representative of landmark regions. Thus,

the SIFT feature points far away from those midpoints

are discarded as they might be incorrect matches. Due

to the lighting condition or other imaging issues of

the outdoor scenes, occasionally the line features are

not very reliable and the true SIFT matches are not

close to the midpoints of line segments. To deal with

this problem, we check the number of SIFT matches

which are close to the midpoints. If the proportion is

less than 1/3, then the line features will not be used to

ﬁlter out the SIFT matching points.

2.3 Camera Heading and Landmark

Information

In the previous section, the most correct SIFT

matches are recorded. The k-means algorithm is then

utilized to cluster the SIFT features, and the landmark

in the panoramic image is identiﬁed by the cluster

with the most SIFT points. If the perspective image is

captured when facing the landmark, then the cluster’s

location in the panoramic image can be used to deter-

mine the camera heading. We convert the heading to

an angle D

between 0 and 360 degrees by

360

− P

) (1)

where P

is the x-coordinate of the cluster’s center in

the panoramic image, P

is the north direction in the

images we record in the database, and L

is the image

width.

The system will collect all of the landmark mark-

ers we have constructed on Google Map if the dis-

tance is less than 90 meters to the input image loca-

tion, and and record all angles D

between the input

location and the landmarks (see Figure 5), i.e.,

= {D

, D

, . . . , D

}

where n is the number of landmark. We choose a

marker M

with the property that the angle difference

between D

and D

is the smallest and given by

= argmin|D

− D

The landmark information can then be derived from

the associated marker on Google Map.

3 EXPERIMENTS

Our visual navigation assistance technique is evalu-

ated on 4 databases captured with a Point Grey La-

dybug2 omnidirectional camera and input images ac-

quired from a smart phone. The geographic infor-

mation of the database and input images are obtained

Figure 5: The camera heading angle between the current

location and the marker.

VISAPP 2016 - International Conference on Computer Vision Theory and Applications

588

Figure 6: An example of the panoramic image in the

database. Top: the original image. Bottom: the cropped

image and used for landmark identiﬁcation.

from a GPS receiver and the built-in GPS module of

the mobile phone, respectively. The Ladybug2 omni-

directional imaging system consists of 6 wide angle

cameras. Six individual images captured by Lady-

bug2 are stitched to form a single panoramic image

with the resolution of 2048 × 1024. Since the lower

part of street scene panoramic images is less impor-

tant for landmark detection, the images are further

cropped to the resolution of 2048 × 598 with the up-

per part remained (see Figure 6). The input image

captured by the mobile phone has the resolution of

3624 × 2448. To reduce the computation time, the

images are resampled to the resolution of 1088 × 816

for processing.

Four image databases used in the experiments are

constructed with different geographic regions. The

detail information is shown in the Table 1 with the to-

tal distance, the numbers of images, nodes and land-

marks. The navigation system graphical user inter-

face is shown in Figure 7. Given an input perspec-

tive image acquired by a mobile camera, it will return

the most similar panoramic image in the database and

draw a red circle as the camera heading derived from

the image. The circle location is P

as used in Eq. (1).

In the navigation system, the name of the landmark

is also provided according to the marker identiﬁed on

Google Map.

Table 1: The numbers of node, image and landmark ob-

tained from the datasets.

Set 1 Set 2 Set 3 Set 4

Distance (m) 673 375 1060 286

# of Node 27 14 41 12

# of Image 29 19 51 15

# of Landmark 4 5 9 3

3.1 Runtime Discussion

We test our navigation assistance system on a Win-

dows 7, 32bit PC with 4GB RAM and an Intel Core

i5 CPU. The run time of the proposed technique is

tabulated in Table 2 with processing stages. The total

execution time for all processes is 8 seconds per im-

age. Due to the image retrieval stage using 10 high

resolution panoramic images (2048 × 598), it spends

most execution time in the process and make it the

bottleneck of our system.

Table 2: The computation time for each processing stage

and the overall execution time.

Process Time (s)

Image retrieval 4.14

Extract SIFT features (query) 0.43

Extract SIFT features (dataset) 0.48

Extract line feature (query) 0.38

Extract SIFT feature (dataset) 0.47

Total execution time 8

3.2 Experimental Results

We ﬁrst compare the landmark matching results. In

our system, we use line features to ﬁlter out the incor-

rect SIFT matching points. As illustrated in Figure

8, the matching results of the input and panoramic

images are shown with different methods. The top

image in Figure 8(b) shows the result only using the

SIFT features to match the input image. The small red

circles represent the matching points. It can be seen

that there are some incorrect SIFT matches which

do not correspond to the landmark’s nearby regions

in Figure 8(a). If the line features are incorporated

as shown in the bottom of Figure 8(b), most of the

wrong SIFT correspondences are eliminated success-

fully. The camera heading of Figure 8(a) is then iden-

tiﬁed correctly.

Figure 7: The graphical user interface of the visual naviga-

tion assistance system.

Visual Navigation with Street View Image Matching

589

(a)

(b)

Figure 8: The results of two landmark matching approaches. The right top result only adopt SIFT to match the landmark. The

right bottom result uses the additional line feature.

To evaluate the system performance, two impor-

tant criteria for navigation assistance are adopted,

namely the correctness of landmark’s name and cam-

era’s heading direction. The camera heading is drawn

in the user interface and indicated by a red circle (say,

). Its correctness is veriﬁed by the overlap with the

landmark area (say, P

) in the panoramic image, i.e.,

∩ P

. Our technique is compared with the meth-

ods proposed by Guan et at. (Guan et al., 2014) and

Bettadapura et al. (Bettadapura et al., 2015) for per-

formance evaluation. Due to the objectives of their

works are not completely the same as ours. Only the

part of their frameworks related to image retrieval and

navigation results is used for comparison. The system

accuracy is computed using our formulation.

The image features derived from their methods are

adopted to calculate the camera heading using Eq. (1),

and the landmark’s name is obtained from the cor-

responding marker in Google Map. In (Guan et al.,

2014), the geographic information and SURF features

are adopted for image retrieval on a pre-constructed

image database. Bettadapura et al. use the geo-

graphic information to retrieve the image. The max-

imally stable extremal region (MSER) is extracted to

detect candidate points, and SIFT is used to describe

the feature points. Table 3 shows the accuracy of our

approach and these two methods performed on each

datasets.

Table 3: The performance comparison with other methods.

Set 1 Set 2 Set 3 Set 4

Our method 79.3% 78.9% 96% 80.0 %

Guan 58.6% 63.2% 76.4% 60.0%

Bettadapura 24.5% 24.1% 47.0% 33.3%

The average accuracy of our results on the datasets

is 83.6%, which is better than those obtained from the

other two methods. Some of our false navigation re-

sults are the incorrect camera headings. It is due to

the encoding error on x-coordinates in the panoramic

images. In (Guan et al., 2014) and (Bettadapura et al.,

2015), they both use the gyroscopic compass to im-

prove the performance of their systems. However, our

technique aims to use only the images and geographic

information to construct the visual navigation assis-

tance. Although the gyroscopic compass can provide

more correct camera headings in the previous works,

our system has better matching results in small ob-

jects like monuments or road signs.

4 CONCLUSIONS

In the paper, we propose a visual navigation assis-

tance system using the geographic information and

image matching. Our technique has been tested on

the real scene images and provides 83.55% accuracy

in four image datasets. In addition to the results

are better than the previous approaches, the proposed

method is also capable of landmark recognition with-

out training. The future work will focus on the inte-

gration with Google Map Street View. With the ad-

vances of wireless communication, the Google street

view image database can be accessed in real-time.

The wearable devices will also be tested to emphasize

the portability of the presented technique.

VISAPP 2016 - International Conference on Computer Vision Theory and Applications

590

ACKNOWLEDGMENTS

The support of this work in part by the National Sci-

ence Council of Taiwan under Grant NSC-102-2221-

E-194-019 is gratefully acknowledged.

REFERENCES

Altwaijry, H., Moghimi, M., and Belongie, S. (2014). Rec-

ognizing locations with google glass: A case study. In

Applications of Computer Vision (WACV), 2014 IEEE

Winter Conference on, pages 167–174.

Bay, H., Ess, A., Tuytelaars, T., and Van Gool, L. (2008).

Speeded-up robust features (surf). Comput. Vis. Image

Underst., 110(3):346–359.

Bettadapura, V., Essa, I. A., and Pantofaru, C. (2015). Ego-

centric ﬁeld-of-view localization using ﬁrst-person

point-of-view devices. In 2015 IEEE Winter Con-

ference on Applications of Computer Vision, WACV

2014, Waikoloa, HI, USA, January 5-9, 2015, pages

626–633. IEEE Computer Society.

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

tection. IEEE Trans. Pattern Analysis and Machine

Intelligence, 8(6):679–698.

Chen, D. M., Baatz, G., Koser, K., Tsai, S. S., Vedan-

tham, R., Pylvanainen, T., Roimela, K., Chen, X.,

Bach, J., Pollefeys, M., Girod, B., and Grzeszczuk,

R. (2011). City-scale landmark identiﬁcation on mo-

bile devices. In Proceedings of the 2011 IEEE Con-

ference on Computer Vision and Pattern Recognition,

CVPR ’11, pages 737–744, Washington, DC, USA.

IEEE Computer Society.

Guan, T., Fan, Y., Duan, L., and Yu, J. (2014). On-device

mobile visual location recognition by using panoramic

images and compressed sensing based visual descrip-

tors.

Hough, P. V. (1962). Method and means for recognizing

complex patterns. US Patent 3,069,654.

Hu, R. and Collomosse, J. (2013). A performance eval-

uation of gradient ﬁeld hog descriptor for sketch

based image retrieval. Comput. Vis. Image Underst.,

117(7):790–806.

Huang, W.-T., Tsai, C.-L., and Lin, H.-Y. (2012). Mobile

robot localization using ceiling landmarks and images

captured from an rgb-d camera. In Advanced Intelli-

gent Mechatronics (AIM), 2012 IEEE/ASME Interna-

tional Conference on, pages 855–860.

Hughes, J. F., van Dam, A., McGuire, M., Sklar, D. F.,

Foley, J. D., Feiner, S. K., and Akeley, K. (2013).

Computer graphics: principles and practice (3rd ed.).

Addison-Wesley Professional, Boston, MA, USA.

Juan, L. and Gwun, O. (2009). A comparison of sift, pca-sift

and surf. International Journal of Image Processing

(IJIP), 3(4):143–152.

Laungrungthip, N., McKinnon, A., Churcher, C., and

Unsworth, K. (2008). Edge-based detection of sky

regions in images for solar exposure prediction. In

Image and Vision Computing New Zealand, 2008.

IVCNZ 2008. 23rd International Conference, pages 1–

Lin, H., Lin, Y., and Yao, J. (2013). Scene change de-

tection and topological map construction using om-

nidirectional image sequences. In Proceedings of the

13. IAPR International Conference on Machine Vision

Applications, MVA 2013, Kyoto, Japan, May 20-23,

2013, pages 57–60.

Liu, H., Mei, T., Luo, J., Li, H., and Li, S. (2012). Finding

perfect rendezvous on the go: Accurate mobile visual

localization and its applications to routing. In Pro-

ceedings of the 20th ACM International Conference

on Multimedia, MM ’12, pages 9–18, New York, NY,

USA. ACM.

Lowe, D. G. (2004). Distinctive image features from scale-

invariant keypoints. Int. J. Comput. Vision, 60(2):91–

110.

Mulloni, A., Wagner, D., Barakonyi, I., and Schmalstieg, D.

(2009). Indoor positioning and navigation with cam-

era phones. Pervasive Computing, IEEE, 8(2):22–31.

Philbin, J., Chum, O., Isard, M., Sivic, J., and Zisserman, A.

(2007). Object retrieval with large vocabularies and

fast spatial matching. In IEEE Conference on Com-

puter Vision and Pattern Recognition, pages 1–8.

Resch, B., Lang, J., and Lensch, H. (2014). Local image

feature matching improvements for omnidirectional

camera systems. In Pattern Recognition (ICPR), 2014

22nd International Conference on, pages 918–923.

Saab, S. and Nakad, Z. (2011). A standalone rﬁd indoor po-

sitioning system using passive tags. Industrial Elec-

tronics, IEEE Transactions on, 58(5):1961–1970.

Siddiqui, J. and Khatibi, S. (2014). Semantic urban maps. In

Pattern Recognition (ICPR), 2014 22nd International

Conference on, pages 4050–4055.

Turcot, P. and Lowe, D. (2009). Better matching with

fewer features: The selection of useful features in

large database recognition problems. In Computer Vi-

sion Workshops (ICCV Workshops), 2009 IEEE 12th

International Conference on, pages 2109–2116.

Uchiyama, H., Saito, H., Servieres, M., and Moreau, G.

(2009). Image based view localization system retriev-

ing from a panorama database by SURF. In Proceed-

ings of the IAPR Conference on Machine Vision Ap-

plications (IAPR MVA 2009), Keio University, Yoko-

hama, Japan, May 20-22, 2009, pages 118–121.

Wang, S., Wang, Y., and Zhu, S.-C. (2015). Learning hier-

archical space tiling for scene modeling, parsing and

attribute tagging. Pattern Analysis and Machine Intel-

ligence, IEEE Transactions on, PP(99):1–1.

Yao, C.-W., Cheng, K.-S., and Lin, H.-Y. (2014). A vi-

sion assisted vehicle navigation technique based on

topological map construction and scene recognition.

In Advanced Video and Signal Based Surveillance

(AVSS), 2014 11th IEEE International Conference on,

pages 399–404.

Zamir, A. R. and Shah, M. (2010). Accurate image local-

ization based on google maps street view. In Proceed-

ings of the 11th European Conference on Computer

Vision: Part IV, ECCV’10, pages 255–268, Berlin,

Heidelberg. Springer-Verlag.

Visual Navigation with Street View Image Matching

591