Tactile Tile Detection Integrated with Ground Detection using an

RGB-Depth Sensor

Yutaro Yamanaka

, Eichi Takaya

and Satoshi Kurihara

Faculty of Science and Technology, Keio University, Yokohama, Japan

Graduate School of Science and Technology, Keio University, Yokohama, Japan

Keywords:

Industrial Applications of AI, Tactile Tile, Ground Detection, Visual Impairment.

Abstract:

Tactile paving is a system used to help visually impaired individuals walk safely. However, it is difﬁcult to

recognize the surrounding tactile tiles on a ﬁrst visit to an area. In this study, we propose a method for detecting

tactile tiles integrated with ground detection using an RGB-Depth sensor. For the ground detection, we use

the RANSAC algorithm and expand the region by using the breadth-ﬁrst search. When detecting the tactile

tiles, we perform thresholding and construct a model to identify candidate areas. Experimental results showed

that the proposed method obtained a precision of about 83% in detecting tactile tiles on a paved asphalt road.

It was possible to correctly distinguish tactile tiles from other objects by combining ground detection in many

cases. On the other hand, there were many false detections of tactile tiles in challenging environments, and

the processing speed should be improved for real-time navigation.

1 INTRODUCTION

According to World Health Organization statistics

(World Health Organization, 2019), as of 2019, there

are 2.2 billion people around the world who have

some form of vision impairment or blindness. Gener-

ally, it is risky for visually impaired individuals to go

out alone. Therefore, a guide with professional quali-

ﬁcations is often asked to accompanythem. Neverthe-

less, they sometimes feel overburdened because they

have to worry about both their surroundings and the

labor of hiring the guide. Even when going out to un-

familiar places with a guide, they are concerned about

their surroundings and experience a lot of stress. For

such individuals, a tactile paving system is indispens-

able for walking safely. As they walk, they use a

white cane or the sole of the shoe to recognize the

protrusions on the surface of the tactile paving. This

system is used in many countries due to its useful-

ness in safely guiding visually impaired people. How-

ever, tactile paving has a problem in that the tiles can-

not be recognized unless the individual is standing on

them. Hence, it is difﬁcult to search for surround-

ing tactile tiles in unfamiliar places. Simply installing

these tiles is not enough to ensure the safety of visu-

ally impaired individuals walking alone. Considering

the above, there is a need for a guide system that can

provide the visually impaired with information on the

surrounding environment, including tactile tiles. By

providing information for different environments and

helping with navigation, such a system will help them

go out safely and easily.

When detecting tactile tiles, two features must be

considered: color and shape. As the international

standard (ISO 23599, 2012) deﬁnes, the color of tac-

tile tiles is typically yellow, as this color is easy to

distinguish from paved asphalt roads. As for shape,

the surface of each tile is lined with linear protrusions

for guidance or point-like protrusions for calling at-

tention. A previous work proposed detecting tactile

tiles from images of the sidewalk by means of com-

puter vision algorithms (Ghilardi et al., 2016). Our

work extends this method so that it can be applied in

outdoor environments where multiple objects exist.

In any guide system for the visually impaired,

false detection is a very serious problem. For exam-

ple, if a tactile tile is falsely detected on a wall and

that navigation information is transmitted to the vi-

sually impaired, they will walk in that direction and

may actually collide with the wall. Generally, tac-

tile tiles are installed on ﬂat ground. If a tactile tile

is detected at a location that is not estimated to be

ﬂat ground in the image, the navigation information

should not be transmitted. We propose a method that

detects both tactile tiles and ﬂat ground in parallel by

using an RGB-Depth sensor. Our contributions are

750

Yamanaka, Y., Takaya, E. and Kurihara, S.

Tactile Tile Detection Integrated with Ground Detection using an RGB-Depth Sensor.

DOI: 10.5220/0009092907500757

In Proceedings of the 12th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2020) - Volume 2, pages 750-757

ISBN: 978-989-758-395-7; ISSN: 2184-433X

summarized as follows.

• We detect tactile tiles in outdoor environments

where various objects exist.

• We reduce false detection of tactile tiles compared

to the case of detecting them in an entire image by

detecting ﬂat ground and narrowing the detection

range.

• Processing time can be shortened compared to the

case of serial processing by executing two pro-

cesses in parallel: ground detection and tactile tile

detection.

2 RELATED WORK

Many approaches to the detection of speciﬁc objects

have been proposed for navigation systems used by

the visually impaired. In some approaches, the seg-

mentation of ﬂat ground is performed using an RGB-

Depth sensor. Yang, Wang, Hu, and Bai proposed

a method to estimate normal vectors and extend re-

gions from depth information after detecting the plane

by using the RANSAC algorithm (Yang et al., 2016).

This method makes it possible to perform the navi-

gation by detecting areas where the visually impaired

can pass. The ground detection process in our work

draws on this technique. Caraiman et al. introduced a

method to detect speciﬁc objects such as doors, stairs,

and signs after completing the ground segmentation

(Caraiman et al., 2017).

Other approaches detect the tactile tiles from a

color image. Kassim et al. developed a method to

determine the types of protrusion on the surface of

a tactile tile by deriving the metric that represents

the geometrical characteristics of the ﬁgure (Kassim

et al., 2018). Jie, Xiaochi, and Zhigang detected the

straight lines of a tile by thresholding, edge detection,

and Hough transforming (Jie et al., 2010). Ghilardi,

Macedo, and Manssour proposed a method that de-

tects a tactile tile from sidewalk images using com-

puter vision algorithms and a decision tree (Ghilardi

et al., 2016). Our work extends these ideas for appli-

cation to outdoor environments where there are mul-

tiple objects besides tactile tiles.

Some approaches detect multiple objects simulta-

neously by semantic segmentation using a convolu-

tional neural network. Among these, a representative

example is a method by Yang et al. that detects ob-

jects such as sidewalks, stairs, and cars by means of a

unique model architecture (Yang et al., 2018). When

providing navigation for the visually impaired, it is

crucial to detect multiple objects simultaneously. In

our work, we also ﬁnd semantic segmentation to be

effective for detecting multiple objects. However, we

need to devise additional approaches for objects such

as tactile tiles, where false detection can be a serious

problem.

3 PROPOSED METHOD

The ﬂow chart of the proposed method is shown in

Figure 1. In this section, we describe the RGB-Depth

sensor we use for our experiment (Sec 3.1), ground

detection (Sec 3.2), tactile tile detection (Sec 3.3), and

the comprehensivejudgment for determining whether

tactile tiles do indeed exist (Sec 3.4).

Figure 1: Flow chart of proposed method.

3.1 RGB-Depth Sensor

We use the Intel RealSense Depth Camera D435i as

the RGB-Depth sensor (Figure 2). This is a stereo

vision camera that can measure depth. It is equipped

with two depth sensors, an RGB camera, and an active

IR projector that illuminates the object. It can also

acquire linear acceleration and angular velocity syn-

chronously with depth information, as it is equipped

with an inertial measurement unit (IMU). As shown in

Figure 3(b), there are some parts of the image where

the depth information is not obtained accurately due

to noise and lack of data. In this work, we remove

the noise and compensate for the lack of data by us-

ing the hole-ﬁlling ﬁlter that comes with RealSense

SDK. The effect of this ﬁlter is shown in Figure 3(c).

Figure 2: RealSense D435i.

3.2 Ground Detection

This subsection describes the coordinate conversion

(Sec 3.2.1), ground estimation (Sec 3.2.2), and esti-

mated region expansion (Sec 3.2.3) in detail.

Tactile Tile Detection Integrated with Ground Detection using an RGB-Depth Sensor

751

(a) (b) (c)

Figure 3: (a) Color image acquired by RGB camera (Re-

alSense D435i). (b) Original depth image. (c) Depth image

after applying hole-ﬁlling ﬁlter to (b).

3.2.1 Coordinate Conversion

We use a three-dimensional point cloud calculated

from the depth information to detect the ground. This

point cloud is acquired as coordinates in the camera

coordinate system, so we need to convert it to coor-

dinates in the global coordinate system. When con-

verting the coordinates, a rotation matrix based on

the attitude of the camera is required. We use the

quaternions to express the posture. The quaternions

are a four-dimensional vector that extends the com-

plex number, which is used to express the posture of

an object in 3D space. The quaternion q is shown in

Equation (1), where q

, q

, and q

are real num-

bers, and i, j, and k are the basic quaternion units.

We use the Madgwick ﬁlter (Madgwick, 2010) to de-

rive the quaternions. According to Equation (2), the

point (X, Y, Z) in the camera coordinatesystem is con-

verted into the point (X

, Y

, Z

) in the global coordi-

nate system by using the quaternion q obtained by the

Madgwick ﬁlter (Diebel, 2006). In Equation (2), R is

equivalent to the rotation matrix.

q = q

+ q

i+ q

j+ q

k (1)

R =



+ q

−q

2(q

−q

) 2(q

+ q

)

2(q

+ q

) q

−q

+ q

−q

2(−q

+ q

)

2(q

−q

) 2(q

+ q

) q

−q

+ q











= R









(2)

3.2.2 Ground Estimation

We estimate the ground by using the random sam-

ple consensus (RANSAC) algorithm (Fischler and

Bolles, 1981), a robust estimation algorithm that con-

siders outliers in the given data and suppresses their

effects. RANSAC estimates the model parameters by

dividing a set of data into inliers (a set of data that the

model ﬁts) and outliers (a set of data that the model

does not ﬁt). The plane model of the point cloud in

the global coordinate is expressed in Equation (3) us-

ing four parameters A, B, C, and D. The condition to

be regarded as inliers is deﬁned by Equation (4) and

consists of the threshold T for the distance between

the estimated plane and the point cloud. First, three

points are randomly selected from the point cloud.

Then, the parameter estimation is performed repeat-

edly by RANSAC. (See the work of Zeineldin and

El-Fishawy (Zeineldin and El-Fishawy, 2016) for a

detailed explanation of how the parameters are esti-

mated.) When the angle between the estimated plane

and the XY plane in Equation (5) is abnormally large,

(i.e., not considered to be the ground), we will return

the processing to the beginning of the loop to reduce

calculation time. In this way, ground with a small an-

gle between the XY plane is estimated. Figure 4(b)

depicts the estimated ground region.

+ BY

+CZ

= D (3)

|AX

+ BY

+CZ

+ D|

√

+ B

< T (4)

= arccos

|B|

√

+ B

(5)

3.2.3 Estimated Region Expansion

As shown in Figure 4(b), only a part of the ground

is estimated by RANSAC. Hence, we need to ex-

pand the estimated region. In this work, we do this

by choosing seeds and performing the breadth-ﬁrst

search. Algorithm (1) describes the details. First, sev-

eral points are randomly selected if two conditions are

satisﬁed: one, points belong to inliers in the model

estimation by RANSAC, and two, one or more four-

connected neighbors of them belong to outliers. Pix-

els of the color image corresponding to those points

are adopted as seeds S. If one of those pixels exists at

the outer edge of the color image, it is not adopted as

a seed. Then, the breadth-ﬁrst search starts from the

ﬁrst element of S, and the pixel visited in the search

is regarded as the ground. If a pixel meets at least one

of the following four conditions, we do not search fur-

ther from that pixel.

• The point in the point cloud corresponding to the

visited pixel p belongs to inliers.

• The visited pixel p has already been searched

• The difference between the hue h

of the visited

pixel p and the average hue h

avg

of pixels that

have corresponding points belonging to inliers is

greater than or equal to the threshold H.

• The visited pixel p is located at the Canny edge

(Harris et al., 1988) of the color image.

The search is repeated for all the elements in S.

The result of the estimated region expansion is shown

ICAART 2020 - 12th International Conference on Agents and Artiﬁcial Intelligence

752

in Figure 4(b) and 4(c), which depicts the expansion

of the region that is considered to be the ground.

Algorithm 1 : Region Expansion based on Breadth-First

Search.

Seeds S = [S

, ..., S

]

avg

= average hue value of inliers

for i = 1 to N do

Initialize Queue Q

Add S

to Q

while Q is not empty do

q = the ﬁrst element of Q

Remove q from Q

for p = four-connected neighbors of q do

if p is not in the inliers and p is not visited

and |h

−h

avg

| < H and

p is not at the Canny edge of the image

then

Add p to Q

p ← visited

end if

end for

end while

end for

(a) (b) (c)

Figure 4: (a) Color image acquired by RGB camera (Re-

alSense D435i). (b) Region estimated to be ground by

RANSAC (pink). (c) Region added to (b) by expansion

(purple).

3.3 Tactile Tile Detection

This subsection describes the thresholding (Sec

3.3.1), noise reduction by DBSCAN clustering (Sec

3.3.2), and model construction by extracting features

(Sec 3.3.3) in detail.

3.3.1 Thresholding

From the viewpoint of visibility, the color of tactile

tiles is usually yellow. First, we convert the RGB im-

age into an image in the YCbCr color space, which

is robust to light conditions, as a previous work (Ghi-

lardi et al., 2016) showed. The YCbCr color space

expresses the color with the luminance (Y), the blue-

difference chroma (Cb) obtained by subtracting the

luminance from blue, and the red-difference chroma

(Cr) obtained by subtracting the luminance from red.

Thus, we can perform thresholding regardless of the

brightness of the image because the hue and bright-

ness are independent in the YCbCr color space. Sec-

ond, we create a histogram for Cb and Cr and perform

thresholding. Figure 5(b) and (e) shows the threshold-

ing results.

3.3.2 Noise Reduction

When candidate areas are detected by thresholding, a

small area is sometimes found, as indicated in Figure

5(b) and (e). Small areas like this are considered a

noise in the detection of tactile tiles because they ex-

ist in a large area of candidate areas. We reduce the

noise with DBSCAN clustering (Ester et al., 1996)

in the pixel coordinate system with the upper-left of

the image as the origin. Since DBSCAN clustering

is a density-based clustering method, which makes it

unnecessary to determine the number of clusters in

advance, it is robust against outliers. In the result of

DBSCAN clustering against remaining pixels as can-

didate areas, some pixels do not belong to any clus-

ters. These pixels are considered noise and are re-

moved. Figure 5(c) and (f) shows the result of noise

reduction. Identiﬁcation after the noise reduction is

necessary because multiple clusters sometimes exist

as candidate areas in outdoor environments.

(a) (b) (c)

(d) (e) (f)

Figure 5: (a), (d) Color image acquired by RGB camera

(RealSense D435i). (b), (e) Color image including noise af-

ter thresholding. (c), (f) Denoised color image by DBSCAN

clustering, where one cluster in (c) and three clusters in (f)

exist as candidates areas.

3.3.3 Model Construction

The shape of the tactile paving surface is unique com-

pared with other objects in an outdoor environment.

Constructing a model that captures this feature is key

for determining whether or not a tactile tile exists in

a candidate area. We use a random forest classiﬁer

(Breiman, 2001) based on a decision tree as the model

to judge its existence. The image texture features

Tactile Tile Detection Integrated with Ground Detection using an RGB-Depth Sensor

753

calculated from the gray-level co-occurrence matrix

(GLCM) are adopted as features of the model. First,

the GLCM is calculated for each offset deﬁned by dif-

ferent distances and angles. Second, the image texture

features are acquired by extracting statistic informa-

tion from these GLCMs. In this work, we obtain them

through the following procedure.

1. Several pixels are chosen randomly from near the

center of one candidate area.

2. Twenty-ﬁve-pixel square images are cropped

from the original image (Figure 5(a) and (d))

around these pixels. These images are regarded

as patches

3. Steps 1 and 2 are repeated for each candidate area.

4. For each patch, the image texture features are ob-

tained by calculating the GLCM with the deter-

mined offsets.

Pixels are selected from near the center of candi-

date areas in step 1 so that we can create patches that

capture the major part of the candidate area. We use

three different distances (1, 2, and 3 pixels) and four

different angles (0, 45, 90, and 135 degrees) as an off-

set and extract six different statistic information, so

a total of 72 features are generated from each patch.

Then, we label these patches as to whether they con-

tain tactile tiles or not. Lastly, the learning of the

model is initiated.

3.4 Comprehensive Judgment

The calculation time to process ground detection and

tactile tile detection in series is long. If these two pro-

cesses are executed in parallel, we can reduce the cal-

culation time. In the result of parallel processing, we

obtain outputs of the ground detection and tactile tile

detection separately. Hence, it is necessary to make

a comprehensive judgment as to whether the tactile

tile does exist. In this study, we count the number of

patches if the following two conditions are satisﬁed.

• The model (Sec 3.3.3) prediction of the patch is

positive.

• The center of the patch is considered to be the

ground in the ground detection (Sec 3.2).

If this number is equal to or greater than a certain

percentage (P) of the number of patches in the area,

the candidate area is considered to be a tactile tile.

4 EXPERIMENTS

This section gives an overview of the experiments we

performed to evaluate the proposed method (Sec 4.1)

and reports the results for the effectiveness of incor-

porating ground detection processing (Sec 4.2), the

parallelism of the proposed method (Sec 4.3), and the

robustness to environments (Sec 4.4).

4.1 Overview of Experiment

To evaluate the proposed method, we performed the

following three experiments.

• Evaluation of the effectiveness of incorporating

ground detection (Sec 4.2).

• Evaluation of the parallelism of the proposed

method (Sec 4.3).

• Evaluation of the robustness to environments (Sec

4.4).

We held the RGB-Depth sensor at a height of

about 1.5 m from the ground surface using a hand-

held tripod. We tilted it from the horizontal at an an-

gle of about 45 degrees toward the ground. Figure

6 shows the experimental scene. We performed the

processing of the proposed method by a laptop com-

puter with an Intel Core i7 processor and 16GB of

memory. The parameter values are listed in Table 1

and we created ﬁve patches (See Sec 3.3.3) per candi-

date area. The identiﬁcation results of candidate areas

are labeled as True Positive (TP), True Negative(TN),

False Positive (FP), or False Negative (FN).

Regarding an area that is not a tactile tile as pos-

itive (FP) is a serious problem. Therefore, the pre-

cision shown in Equation (6) is used to evaluate the

identiﬁcation results.

Figure 6: Experimental scene in outdoor environment.

Table 1: Parameter values used in experiment.

T (Sec 3.2.2) 0.03 m

(Sec 3.2.2) 10

◦

N (Sec 3.2.3) 50

H (Sec 3.2.3) 20

P (Sec 3.4) 40 %

Precision =

TP+ FP

(6)

ICAART 2020 - 12th International Conference on Agents and Artiﬁcial Intelligence

754

4.2 Effectiveness of Incorporating

Ground Detection

4.2.1 Experimental Design

To evaluate the effectiveness of incorporating ground

detection, we took videos with the RGB-Depth sensor

in the following two environments.

• Tactile tiles on a paved asphalt road in a sunny

area (measuring distance: about 200 m) (Env 1).

• Tactile tiles on a paved asphalt in a shaded area

(measuring distance: about 200 m) (Env ).

When identifying candidate areas in each frame,

we performed two operations simultaneously: nar-

rowing the detection range of tactile tiles incorporat-

ing ground detection, and detecting them from the en-

tire image. We examined the following four cases.

1. Detecting the ground and narrowing the detection

range of tactile tiles in Env 1 (Case 1).

2. Detecting tactile tiles from the entire image with-

out ground detection in Env 1 (Case 2).

3. Detecting the ground and narrowing the detection

range of tactile tiles in Env 2 (Case 3).

4. Detecting tactile tiles from the entire image with-

out ground detection in Env 2 (Case 4).

In Cases 2 and 4, only the ﬁrst condition of the

comprehensive judgment (Sec 3.4) was valid. We cal-

culated the precision of the identiﬁcation results and

compared them.

4.2.2 Results and Discussion

Table 2 lists the results of the candidate area iden-

tiﬁcation for each of the four cases. As shown, the

precision was higher when ground detection was per-

formed in both Env 1 and Env 2. We conclude that

incorporating ground detection is effective for reduc-

ing false detections of tactile tiles.

Table 2: Effectiveness of ground detection.

Case TP FP FN TN Precision

1 171 31 155 164 0.8465

2 206 73 120 122 0.7384

3 95 20 183 398 0.8261

4 162 124 116 294 0.5664

4.3 Parallelism of Proposed Method

4.3.1 Experimental Design

In the proposed method, ground detection and tactile

tile detection are performed in parallel. To evaluate

the parallelism, we performed an experiment where

two processes are performed in series for comparison.

First, we excluded pixels that were not considered

ground in the ground detection (Sec 3.2). Second, we

performed tactile tile detection (Sec 3.3) on the re-

maining pixels. As a result, two processes were per-

formed sequentially. We took videos with the RGB-

Depth sensor in Env 1. For these two cases, we calcu-

lated the average processing time per frame and com-

pared it.

4.3.2 Results and Discussion

Table 3 lists the average processing time per frame for

these two experiments. As shown, the processing time

per frame can be shortened by processing in parallel.

Table 3: Evaluation of parallel processing.

Type Processing time per frame

Serial 1.5904 sec

Parallel 1.4832 sec

4.4 Robustness to Environments

4.4.1 Experimental Design

We investigated whether the proposed method can be

applied in various environments. We took videos with

the RGB-Depth sensor in Env 1, Env 2, and the fol-

lowing three environments.

• Tactile tiles with a small luminance ratio with the

road (measuring distance: about 80 m) (Env ).

• Tactile tiles on a tiled sidewalk (measuring dis-

tance: about 80 m) (Env 4).

• A tiled sidewalk with no tactile tiles (measuring

distance: about 80 m) (Env 5).

For each environment, we calculated the precision

of identiﬁcation results and compared them.

4.4.2 Results and Discussion

Table 4 lists the identiﬁcation results of candidate ar-

eas in various environments. Figure 7 shows the re-

sults of tactile tile detection by the proposed method,

where green and blue boxes indicate positive and neg-

ative patches by comprehensive judgment (Sec 3.4),

respectively. As Figure 7(a)-(c) shows, the candidate

Tactile Tile Detection Integrated with Ground Detection using an RGB-Depth Sensor

755

area was identiﬁed relatively accurately, and the pre-

cision was high: 84.7% in Env 1 and 82.6% in Env 2.

In contrast, as Figure 7(d)-(f) shows, there were many

false detections of tactile tiles in Env 3, Env 4, and

Env 5, and the precision was low.

As mentioned in the international standard (ISO

23599, 2012), the tactile tile should have a high lu-

minance ratio with the road. The tactile tile in Env 3

did not follow this standard. It seems that these tac-

tile tiles were not detected correctly because the pro-

posed method expected them to follow the standard.

In Env 4 and Env 5, we presume that tiled sidewalks

were the problem; speciﬁcally, the shape of the tiles

on the sidewalks was similar to the surface of the tac-

tile tiles for guidance. We also think that most of the

pixels remained due to the tile color after threshold-

ing. Furthermore, the model prediction probably gave

false positives because the training data lacked data

that had negative labels of the tiles on the sidewalks.

Table 4: Identiﬁcation results in each environment.

Env TP FP FN TN Precision

1 171 31 155 164 0.8465

2 95 20 183 398 0.8261

3 20 37 59 12 0.3509

4 22 37 12 95 0.3729

5 0 39 0 47 0

(a) (b) (c)

(d) (e) (f)

Figure 7: Detection results in each environment. (a), (b)

Env 1. (c) Env 2. (d) Env 3. (e) Env 4. (f) Env 5.

5 CONCLUSIONS

In this paper, we have proposed a method of detect-

ing ground and tactile tiles in parallel by means of

an RGB-Depth sensor to provide information on the

surrounding tactile tiles to the visually impaired. Ex-

perimental results showed that the proposed method

obtained the precision of about 83% on a paved as-

phalt road.

In future work, we aim to improve the detection

performance of the model and the real-time perfor-

mance of the processing. Moreover, it is vital to de-

vise an actual navigation method. To improve the de-

tection performance, we will use pictures of tactile

tiles in various environments as training data for the

model. Also, although we shortened the processing

time by parallel processing in this study, the speed

was insufﬁcient for real-time navigation. Therefore,

we will optimize the processing or consider another

faster method. Lastly, the information derived from

the detection results should be conveyed to the visu-

ally impaired through voice and so on. This required

information includes the direction and distance of tac-

tile tiles from the current standing point. For the dis-

tance, the depth information acquired by the RGB-

Depth sensor should be useful. We will examine a

concrete navigation method and develop a prototype

that incorporates an RGB-Depth sensor.

REFERENCES

Breiman, L. (2001). Random forests. Machine learning,

45(1):5–32.

Caraiman, S., Morar, A., Owczarek, M., Burlacu, A.,

Rzeszotarski, D., Botezatu, N., Herghelegiu, P.,

Moldoveanu, F., Strumillo, P., and Moldoveanu, A.

(2017). Computer vision for the visually impaired: the

sound of vision system. In Proceedings of the IEEE

International Conference on Computer Vision, pages

1480–1489.

Diebel, J. (2006). Representing attitude: Euler angles,

unit quaternions, and rotation vectors. Matrix, 58(15-

16):1–35.

Ester, M., Kriegel, H.-P., Sander, J., Xu, X., et al. (1996).

A density-based algorithm for discovering clusters in

large spatial databases with noise. In Kdd, volume 96,

pages 226–231.

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.

Communications of the ACM, 24(6):381–395.

Ghilardi, M. C., Macedo, R. C., and Manssour, I. H. (2016).

A new approach for automatic detection of tactile

paving surfaces in sidewalks. Procedia computer sci-

ence, 80:662–672.

Harris, C. G., Stephens, M., et al. (1988). A combined cor-

ner and edge detector. In Alvey vision conference, vol-

ume 15, pages 10–5244. Citeseer.

ISO 23599 (2012). Assistive products for blind and vision-

impaired persons - Tactile walking surface indicators.

Jie, X., Xiaochi, W., and Zhigang, F. (2010). Research

and implementation of blind sidewalk detection in

portable eta system. In 2010 International Forum on

Information Technology and Applications, volume 2,

pages 431–434. IEEE.

ICAART 2020 - 12th International Conference on Agents and Artiﬁcial Intelligence

756

Kassim, A. B. M., Yasuno, T., Suzuki, H., Aras, M. S. M.,

Shukor, A. Z., Jaafar, H. I., and Jafar, F. A. (2018).

Vision-based tactile paving detection method in navi-

gation systems for visually impaired persons. In Ad-

vances in Human and Machine Navigation Systems.

IntechOpen.

Madgwick, S. (2010). An efﬁcient orientation ﬁlter for in-

ertial and inertial/magnetic sensor arrays. Report x-io

and University of Bristol (UK), 25:113–118.

World Health Organization (2019). Fact sheet blindness

and vision impairment. https://www.who.int/news-

room/fact-sheets/detail/blindness-and-visual-

impairment.

Yang, K., Wang, K., Bergasa, L., Romera, E., Hu, W., Sun,

D., Sun, J., Cheng, R., Chen, T., and L´opez, E. (2018).

Unifying terrain awareness for the visually impaired

through real-time semantic segmentation. Sensors,

18(5):1506.

Yang, K., Wang, K., Hu, W., and Bai, J. (2016). Expanding

the detection of traversable area with realsense for the

visually impaired. Sensors, 16(11):1954.

Zeineldin, R. A. and El-Fishawy, N. A. (2016). Fast and

accurate ground plane detection for the visually im-

paired from 3d organized point clouds. In 2016 SAI

Computing Conference (SAI), pages 373–379. IEEE.

Tactile Tile Detection Integrated with Ground Detection using an RGB-Depth Sensor

757