Outdoor Context Awareness Device That Enables Mobile Phone

Users to Walk Safely through Urban Intersections

Jihye Hwang

, Younggwang Ji

, Nojun Kwak

and Eun Yi Kim

Graduate School of Convergence Science and Technology, Seoul National University, Seoul, Korea

Internet & Multimedia Engineering, Konkuk University, Seoul, Korea

Keywords: Guidance System, Mobile Phone User, Outdoor Context Awareness, Support Vector Machine, Multi-scale

Classification.

Abstract: Research in social science has shown that the mobile phone users pay less attention to their surroundings,

which exposes them to various hazards such as collisions with vehicles than other pedestrians. In this paper,

we propose a novel handheld device that assists mobile phone users to walk more safely outdoors. The

proposed system is implemented on a smart phone and uses its back camera to detect the current outdoor

context, e.g. traffic intersections, roadways, and sidewalks, finally alerts the user of unsafe situations using

sound and vibration from the phone. The outdoor context awareness is performed by three steps: pre-

processing, feature extraction, and context recognition. First, it improves the image contrast while removing

image noise, and then it extracts the color and texture descriptors from each pixel. Next, each pixel is

classified as an intersection, sidewalk, or roadway using a support vector machine-based classifier. Then, to

support the real-time performance on the smart phone, a multi-scale classification is applied to input image,

where the coarse layer first discriminates the boundary pixels from the background and the fine layer

categorizes the boundary pixels as sidewalk, roadway, or intersection. In order to demonstrate the

effectiveness of the proposed method, some real-world experiments were performed, then the results

showed that the proposed system has the accuracy of above 98% at the various environments.

1 INTRODUCTION

When walking through an urban area, we can easily

observe pedestrians talking or reading and sending

text messages on mobile phones. A number of

studies on pedestrians crossing streets have shown

that mobile phone users exhibit less safe behavior

than other pedestrians, as their attention has been

directed away from the external environment and

they do not see the overall context (Institute of

transportation engineers, 2004; Kate, 2010; Leena,

2012; Mark, 2009; Rovert, 2010; Tianyu, 2012 ).

The mobile phone users pay less attention to

their surroundings, which exposes them to various

hazards such as collisions with people, cars, and

other obstacles. Among the numerous hazards

outside, traffic intersections are the most dangerous

areas for pedestrians. Approximately 20.3% of all

traffic accidents occur at intersections and this

pedestrian accident rate is one of the highest rates

for all traffic accidents (Jack, 2007; Julie,2006).

During last decades, a number of solutions for

safety crossing the traffic intersections has been

proposed and implemented (Brabyn, 1933; Barlow,

2003; Dragan, 2011; Karacs, 2006; Volodymyr,

2008; Chen-Fu, 2012; James, 2013). Then, the initial

research has been focusing the development of the

safety system for the people with visual

impairments. For example, the audible pedestrian

signals (Barlow, 2003) and talking signs (Brabyn,

1933) have been developed in order to inform the

visually impaired pedestrians to know when to cross

intersections. However, although these solutions are

being adopted more widely, they are still only

available in limited places and additional devices

should be installed or used.

As an alternative to such systems, some vision-

based systems such as Crosswatch (Volodymyr,

2008), Zebralocalizer (Dragan, 2011) and Walksafe

(Tianyu, 2012) have been proposed to prevent the

pedestrians of the collisions with vehicles at urban

traffics. Crosswatch is a handheld vision system that

provides real-time feedback to the user about their

526

Hwang, J., Ji, Y., Kwak, N. and Kim, E.

Outdoor Context Awareness Device That Enables Mobile Phone Users to Walk Safely through Urban Intersections.

DOI: 10.5220/0005664705260533

In Proceedings of the 5th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2016), pages 526-533

ISBN: 978-989-758-173-1

orientation and location at the crosswalk; then, it can

detect the crosswalk using edge information.

Similarly, the Zebralocalizer identifies pedestrian

crossings (i.e. zebra crossings) using line analyses

and localization of crosswalks using a camera and

3D accelerometers. These vision-based methods

function successfully on localizing crosswalks at

intersections and guiding users to cross intersections

safely.

Recently, some commercial products such as

“Text and Walk,” and “Walk and Email,” which

displays the road situation captured from back

camera on application background, thereby making

pedestrians users write SMS and e-mail while

walking safety. However, it was shown in (Ophir,

2009) that the users may not be aware of dangers

even if they are displayed as application

background. In (Sivaraman, 2010), authors

presented a car detection system based on Harr

features that exploits the back camera of

smartphone. It was well worked on the resource

constrained smartphone, however, it can only detect

cars when they were very close to the pedestrians,

limiting the time for the pedestrians to react safety.

On the other hand, none of them have considered

the current situations where a user stands on. In real

scenarios, the people require different guidance

solutions according to the context where they are

currently located. For example, if a user is at an

intersection, they want to locate the crosswalk.

However, when the user is walking on the sidewalk,

the system leads the user to walk on the far side

from the road. Accordingly, the outdoor contexts

where the user is located should be first recognized.

In this paper, a novel method for automatically

recognizing a user’s current context is proposed in

order to increase pedestrian safety, particularly for

users who operate their mobile phone while walking.

Here, the context refers to the type of place where a

user is standing, which is classified as a sidewalk,

roadway, or intersection. Among these types of

contexts, the discrimination between a sidewalk and

an intersection are more important than recognizing

a roadway.

As a key in discriminating outdoor contexts, the

orientation of the boundaries between sidewalks and

roadways are used: horizontally oriented boundaries

are found in images corresponding to intersections

and more vertically oriented boundaries are

observed in images corresponding to sidewalks.

Therefore, localizing such boundaries from input

images should be undertaken first. In order to

separate the boundaries between sidewalks and

roadways from other lines, and then to discriminate

such boundaries as sidewalks or intersections, the

color and texture properties of the images are

considered and machine-learning based

classifications such as a support vector machine

(SVM) are used. Then, in order to improve the

computation cost and accuracy, a multi-scale

classification is adopted, where a coarse layer first

classifies the boundary pixels from the background

and a fine layer classifies the boundary pixels into

one of the three contexts: sidewalks, intersections, or

roadways.

In order to evaluate the effectiveness of the

proposed system, numerous videos were collected

from real environments, and they were used to

measure the accuracy of the proposed system. From

the experimental results, it was found that the

average accuracy was 98.25%.

2 SYSTEM ARCHITECTURE

We propose a novel assistive device that aids mobile

phone users walking and crossing roads more safety.

The proposed system is implemented on smartphone

and uses its back camera to detect users’ current

context, and notifies the recognized results to users

through sound and vibration from the phone.

Then, as a key element of discriminating

between sidewalks and intersections, the orientation

of the boundaries between the sidewalks and

roadways are used. Fig 1 illustrates some sample

intersections and sidewalks captured from outdoors,

where the vertical and horizontal lines in the

boundaries between sidewalks and roadways can be

easily observed. The images corresponding to

intersections have horizontal boundaries, as shown

in Fig. 1(a), whereas the images corresponding to

sidewalks have some boundaries that are close to

vertical and non-horizontal lines (see Fig. 1(b)).

(a) (b)

Figure 1: Some of outdoor images (a) images categorized

to intersections (b) images categorized to sidewalks.

Based on these observations, the proposed

method was designed and developed. As seen in Fig.

1, it is critical to accurately localize the boundaries

from the input images. For this, we use the color and

textural properties, which are trained by machine

learning algorithm.

Outdoor Context Awareness Device That Enables Mobile Phone Users to Walk Safely through Urban Intersections

527

Figure 2: System overview.

The core algorithm to discriminate users’ current

contexts as sidewalk, roadway and intersections is

based on computer vision techniques. Computer

vision processing is a computational intensive

process that can easily drain the computational

resources and batteries of smartphone. To address

this, the proposed system use the simple image

features such as colors and textures, and use the

learning-based recognition model that is first trained

offline and then uploaded and used for the outdoor

situation recognition, as shown in Fig 2.

To build a learning based recognition model (the

right side of Fig 2), we first prepare a dataset

containing positive and negative images, for

example, images that show the sidewalk, roadway

and intersection (see Fig 1). Such sets of images are

first preprocessed to enhance the contrast while

filtering some noises, then input to an algorithm that

extracts characterizing features such as colors and

textures and then used to build classifiers able to

determine if a picture contains sidewalk or traffic

intersection. The resulting classifiers are then used

by the smartphone application running on

smartphones, running the online context recognition

in real-time.

In on-line scenarios, the proposed system is

performed by three steps: preprocessing, feature

extraction, and context recognition, which are

described in the left side of Fig 2. In preprocessing,

some noise and contrast are removed and improved

using Gaussian smoothing and histogram

equalization. Then, the useful visual features are

extracted including the color and textures; saturation

and intensity are used as color descriptors and a

histogram of oriented gradient (HOG) is used to

describe the textural properties. These features then

become the input for the context recognition

module, which separates the boundaries from the

background and re-categorizes them as pixels

corresponding to sidewalks or intersections. In this

study, multi-scale classification is used to reduce the

computational cost of the classifier and improve the

accuracy.

3 FEATURE EXTRACTION

As a key element of discriminating the outdoor

contexts, the orientation of the boundaries between

sidewalks and roadways are used. In order to

discriminate the boundaries between sidewalks and

roadways from other lines, the visual characteristics

such as color and texture are investigated. Here, the

saturation and intensity are used to describe the

color and a histogram of oriented gradient (HOG) is

used to describe the texture.

The color information is used as inputs for

classification problems. In this study, the RGB

image is converted to the HSI color model using

below equation, because the HSI color model

represents colors in a similar manner to that which

human eyes sense them.

Generally, the pixels that correspond to the

roadway have distinctive color information, i.e.

these regions have lower color saturations than

others, as half of the color belongs to roadways that

are marked with gray. Accordingly, the saturation

and intensity are used to discriminate these contexts.

Then, the saturation and intensity for every M×M

sized sub-region is computed as the average of each

quarter sub-region.

In order to describe the textural properties of

sidewalks and intersections, various texture

descriptors have been considered such as the

histogram of oriented gradient (HOG), fast Fourier

transform (FFT) coefficients, and wavelet transform.

Through the experiment, the HOG method was

selected as the most reliable method to describe the

textural properties.

The HOG is a feature descriptor that counts the

occurrences of a gradient orientation in the sub-

regions of an image, and it is often used for object

detection. This method is computed on a dense grid

of uniformly spaced cells and uses overlapping local

contrast normalization for improved accuracy.

In this paper, the contexts are defined as

sidewalks and intersections; these two contexts have

different characteristics in terms of the gradient

orientations. The boundaries of sidewalks are

oriented close to vertical, and the horizontally

oriented boundaries were found in the images

corresponding to intersections. However, other

image parts corresponding to the same objects have

a relatively uniform gradient orientation. Therefore,

ICPRAM 2016 - International Conference on Pattern Recognition Applications and Methods

528

the HOG is effective in both separating the

boundaries between sidewalks and roadways from

other images, and categorizing the contexts.

For all M×M-sized sub-regions, the HOGs are

calculated using the following procedures. First, the

gradient magnitude and orientation of each pixel are

computed using the Sobel operator; then, each

pixel’s gradient orientation is discretized into six

histogram bins, and the pixels’ magnitudes are

accumulated in the corresponding HOG bin. Finally,

the HOG is normalized.

Fig 3 presents HOG characteristics for some

different types of regions. Fig 3(a) shows the HOG

distributions for the sub-regions corresponding to

the sidewalks and intersections, and Fig 3(c) shows

those for parts of some objects. Interestingly, in such

figures, distinctive distributions are seen according

to their region type. Generally, for the sub-regions of

the sidewalks and intersections, the distribution of

the HOGs tends to be biased at specific histogram

bins. However, other regions have relatively

uniformly distributed HOGs. In other words, the first

two histograms have larger variances than the last

one.

(a) (b) (c)

Figure 3: Examples of HOGs (a) and (b) HOGs

distributions of sub-regions corresponding to the

sidewalks and intersections, (c) HOGs distributions of

other sub-regions. (row: HOG’s bin).

4 CONTEXT RECOGNITION

In this module, the current context where a user

stands is recognized based on the colors and

textures. As mentioned above, the key elements for

context recognition are the boundary orientations

between sidewalks and roadways: horizontally

oriented boundaries are found in images

corresponding to intersections and more vertically

oriented boundaries are observed in images

corresponding to sidewalks. Accordingly, the

boundary should be first discriminated from other

natural lines; then, these pixels should be classified.

In the proposed method, such classifications are

accomplished using a multi-scale classification,

which is illustrated in Fig 4.

Figure 4: The multi-scale classification scheme.

The multi-scale classification applies the same

operator to the input image while changing their

scale and this method has been used in numerous

applications including image segmentation and

classification (Zoltan, 1996). The proposed multi-

scale classification mechanism is composed of a

coarse layer and a fine layer: the coarse layer

discriminates the boundary pixels and the fine layer

categorizes the pixels into sidewalk and intersection

pixels. Then, in order to accomplish the different

goals of the respective layers, the SVM is used as a

classifier.

4.1 Locating the Boundary Pixel in the

Coarse Layer

Generally, a boundary is detected using the edge

operator and Hough transform. However, the

boundaries detected using these methods include

many sections that correspond to natural lines as

well as boundary pixels, which increases the

computation costs for the classification process. In

the coarse layer, each boundary pixel is classified as

a boundary pixel from other regions.

As shown in Fig. 4, the original image is first

downscaled by 20×20, thus the average value of

input pixels within a 20×20 block is assigned to the

pixel in the down-sampled image. Then, for each

pixel and its neighbors, the color and texture

properties are extracted.

Thereafter, they input into the SVM-based

classifier. The SVM was originally a binary

classification algorithm developed by Vapnik et al.

and it has been successively extended by numerous

other researchers (Crammer, 2010).

Given a set of training examples that were

classified as boundary or other, the SVM training

algorithm builds a model that assigns new data into

one category or the other. In this paper, the Gaussian

radial basis function (RBF) was used according the

experiment.





,



=











,>0

(1)

123456

Outdoor Context Awareness Device That Enables Mobile Phone Users to Walk Safely through Urban Intersections

529

For the experiment, the parameters were set as

follows: gamma = 1 and cost = 1000.

Fig 5 shows the results of boundary extraction.

For the input image in Fig 5(a), the result is shown

in Fig 5(b), where the boundaries are denoted using

the white color. As can be seen in the figure, the

SVM training algorithm can successfully separate

boundaries from other natural lines.

4.2 Determining the Context of the

Boundary Pixel in the Fine Layer

In the fine layer, the original image sized at 640×480

is used. Accordingly, every boundary pixel expands

to a block with a size of 20×20, each pixel of which

has been categorized as intersection, sidewalk, or

roadway by the SVM-based classifier.

As is well known, the SVM is used to classify

objects into two class types. Accordingly, in order to

to determine the current context from the three

possible contexts, i.e. sidewalk, roadway, or

intersection, two SVMs are used hierarchically, as

follows:





=



2 i

 SVM







1 elsei

SVM









0 else 

(2)

The first SVM classifies a pixel’s class as

sidewalk or other, and then the second SVM

discriminates a pixel’s class as either an intersection

or a roadway. Then, the same visual features that are

used in the coarse layer are exploited.

(a) (b) (c) (d)

Figure 5: Examples of the multi-scale classification

results. (a) input image, (b) boundary extraction results,

post-processing.

Fig 5 presents the outdoor context recognition

results. For the input image in Fig. 5(a), the context

classification result is shown in Fig. 5(c) where the

red pixels and blue pixels are used to denote the

pixel class as a sidewalk or intersection,

respectively. As shown in Fig. 5(c), the

classification results have some errors (noise)

because the classification decision is performed

locally on each pixel. Thus, smoothing is performed

globally on the texture classification results in order

to combine the individual decisions for a whole

image. In order to solve this problem, a grid map

was designed where each cell has a size of 80×80.

The grid map is overlaid on the classification results,

and then the classification decision for each cell is

made. Then, the context of a cell is determined by

the majority rule. For example, if most pixels in a

cell are classified as intersections, the cell is also

considered to be an intersection. Fig 5(d) shows the

final classification result where the red block and

blue block denote the sidewalk and intersection

classes, respectively.

5 EVALUATION

In order to assess the practical validity of the

proposed system, a number of real-world

experiments were performed. The 30-minte walking

experiments were asked to participants from 10:00

to 18:00 hours at six different locations. The

smartphone was held by the participants standing at

the streets. We recorded the captured by a phone,

and counted the number of the respect context that

appeared in the video as ground truth, which were

used for objective evaluation of the proposed

system.

Three females and two males participated in the

evaluation and their average age was 25.6 years. In a

day, all participants used mobile phone 6 hours on

average and usually were used mobile phone on the

road to sending text message, playing the mobile

phone game and so on. In addition, three participants

had the accident experience with person or poles that

stay in front of the intersection while using a mobile

phone on the road.

Each participant was given an introduction on

how to operate the proposed system. Furthermore,

the participants were asked to walk using the

instructions provided by the proposed system. If a

participant arrives at intersections on their way, they

should stop and wait for the traffic signals. Then, we

took participants to unfamiliar traffic intersections to

test the system.

At the selected locations for real-world

experiments, the accuracy of the proposed context

recognition was evaluated. We compared the results

with the ground truth. The average results are shown

in Table 1. The overall accuracy was above 98%.

ICPRAM 2016 - International Conference on Pattern Recognition Applications and Methods

530

Table 1: The accuracy of proposed system.

Location 1 2 3 4 5 6 Total

Accuracy 100 98.4 100 94.4 100 96.7 98.25

Fig 6 some recognition results for various

environments. Fig 6(a) shows the input images,

where the images have time-varying lighting and the

sidewalks have diverse patterns and colors. The

input images were first enhanced in the

preprocessing stage, and then the classifications

were performed.

(a) (b) (c) (d)

Figure 6: Context recognition results. (a) Input image, (b)

boundary extraction results, (c) classification results, (d)

final results after post-processing.

As shown in Fig. 6(b), the boundaries between

the sidewalks and roadways were correctly

extracted, despite the diverse patterns on the

sidewalks; however, they still included some false

classifications. Fig. 6(c) shows the classification

results, and Fig 6(d) shows the final results after

post-processing. Although some cells were

misclassified, the majority of the cells were

classified as intersections; thus, its context was

considered as an intersection. All examples

recognized the correct context even though some

pixels on the boundaries were misclassified. The

results demonstrate that the proposed method has

robust performance in the ground pattern and

illumination type.

Fig 7 presents some examples of errors that

occurred in complex scenes. As seen in Fig. 7, most

boundary sections between the sidewalks and

roadways were concealed by pedestrians and their

shadows, so the boundaries in the fine layer

classification could not be extracted (see the first

row of Fig 7(b)). Thus, two images were

misclassified as sidewalks. However, these errors

can be easily resolved if the history of the place

types between specified time frames are used.

(a) (b) (c) (d)

Figure 7: Examples of misclassified images by moving

walkers. (a) Input image, (b) boundary extraction results,

processing.

The primary purpose of the proposed system is to

increase the safety of pedestrians who use their

mobile phone while walking. For the system’s

practical use as an assistive device, real-time

processing should be supported. The average

processing time was approximately 304.16ms. As

such, the proposed method can process more than

four frames per second on resource-constrained

smartphone. Consequently, the experiments

demonstrated that the proposed method produces

superior accuracy for context awareness, thereby

assisting safer navigation for pedestrians in real-

time.

In order to study the participants’ satisfaction in

using the proposed system, the participant

satisfaction when using the proposed system was

investigated. The following four questions were used

to determine the participant satisfaction.

z E1: How helpful the system when you were

walking?

z E2: How difficult was it to understand the

system?

z E3: Is it comfortable to use the proposed

system in the real context (How much it is

comfortable)?

z E4: Do you want to use this system if it is

developed as smart device app?

Outdoor Context Awareness Device That Enables Mobile Phone Users to Walk Safely through Urban Intersections

531

Fig 8 presents the results of the participant

survey where the middle line represents the standard

deviation. As seen in the figure, most participants

were satisfied with the proposed system.

Figure 8: The user survey results.

In question E1, there was 76% satisfaction with

the proposed system. Participant 5 stated, “I usually

fall down while walking, but if I used this system, I

could prevent a dangerous situation.” Although the

proposed system sometimes recognized incorrect

information, the accuracy for detecting intersections

was 96%, and participants could avoid most

dangerous situations.

Question E2 referred to the ease of

understanding and using the proposed system. The

proposed system sent a warning alarm to the

participant when they were approaching a traffic

intersection and made them stop. Participant 1 said,

“Because I was using the system for the first time, I

was confused when I stopped. However, after

practice I could use the system well without any

effort.” Most participants were slightly confused by

the proposed system when they used it for the first

time, but gradually their skill with and

understanding of the proposed system improved

with use without greater effort.

In question E3, all participants felt that it was

convenient to use the proposed system. The

participants received information that analyzed the

current context using an audio or tactile interface.

Sometimes they complained about the earphones,

because they blocked other outside sounds.

Participant 2 stated, “Usually I use earphones to

listen to music, so I thought that the alert sound

‘beep’ was just noise.”

In question E4, approximately 80% of the

participants agreed that if the system were developed

on a smart phone as multitasking, they would use it.

Participant 5 said, “Actually, now I really need this

system. I always use my smart phone while walking,

and I don’t see intersection quickly enough. So, if

this system was developed as multitasking, I will

download it.” Consequently, in future research, the

proposed system will be developed for multitasking

smart devices.

Consequently, the experiments demonstrated that

the proposed system produces superior accuracy for

context recognition and that participants feel

comfortable in using the proposed system.

6 CONCLUSION

In this study, an assistive device that increases the

safety of the pedestrian, particularly for users who

use their mobile phone while walking, was

presented. The primary goal of the proposed system

is to automatically recognize outdoor contexts such

as intersections, sidewalks, and roadways, thereby

reducing the number of traffic accidents involving

pedestrians and vehicles. The proposed system is

composed of three modules: pre-processing, feature

extraction, and context recognition. In order to

reduce the computational cost and improve the

accuracy, a multi-scale classification was used

where a coarse layer discriminated the boundary

pixels and a fine layer categorized the pixels into

sidewalks and intersections. In both layers, the

support vector machine was used as the classifier.

In order to assess the validity of the proposed

system, real-world experiments were conducted with

five participants. The results demonstrated that the

proposed method could recognize outdoor contexts

with an accuracy of 98.25%, which proved the

practical validity of the proposed system as mobility

aids for pedestrians. To fully support the safety of

pedestrians, obstacle detection and avoidance should

be embedded into the current system. In future

research, this will be extended to the current system.

ACKNOWLEDGEMENTS

This research was supported by Basic Science

Research Program through the National Research

Foundation of Korea(NRF) funded by the Ministry

of Science, ICT & Future Planning(NRF-

2013R1A1A3013445).

REFERENCES

Institute of Transportation Engineers U.S. Department of

Transportation Federal Highway Administration,

100

E 1E 2E 3E 4

Score

Question

User 1 User 2 User 3 User 4 User 5

ICPRAM 2016 - International Conference on Pattern Recognition Applications and Methods

532

2004, Pedestrian safety at intersection,

PEDESTRIANS.

Kate seth, 2010, Walk on the safe side, Traffic Technology

International.

Leena, Rauno, 2012, Accidents and close call situations

connected to the use of mobile phones, Accident

Analysis & Prevention.

Mark B et al., 2009, Pedestrians, vehicles, and cell

phones, Accident Analysis & Prevention.

Robert et al., 2010, Association Between Roadway

Intersection Characteristics and Pedestrian Crash Risk

in Alameda County, California, Transportation

Research Record: Journal of the Transportation

Research Board.

Tianyu et al., 2012, WalkSafe: a pedestrian safety app for

mobile phone users who walk and talk while crossing

roads, ACM HotMobile.

Jack et al., 2007, Mobile telephones, distracted attention,

and pedestrian safety, Accident Analysis & Prevention.

Julie, Susanne, 2006, The effects of mobile phone use on

pedestrian crossing behaviour at signalized and

unsignalized intersections, Accident Analysis &

Prevention.

Brabyn J et al., 1933, Talking signs: a remote signage,

solution for the blind, visually impaired and reading

disabled, Engineering in Medicine and Biology

Society.

Barlow J et al., 2003, Accessible pedestrian signals:

Synthesis and guide to best practice, National

Cooperative Highway Research Program.

Dragan et al., 2011, Zebralocalizer: identification and

localization of pedestrian crossings, ACM MobileHCI.

Karacs K et al., 2006, Bionic Eyeglass: an Audio Guide

for Visually Impaired. Biomedical Circuits and

Systems, BioCAS.

Volodymyr et al., 2008, Crosswatch: a Camera Phone

System for Orienting Visually Impaired Pedestrians at

Traffic Intersections, International Conference on

Computers Helping People with Special Needs.

Chen-Fu Liao, 2012, Using a Smartphone App to Assist

the Visually Impaired at Signalized, University of

Minnesota.

James M et al., 2013, Crosswatch: a System for Providing

Guidance to Visually Impaired Travelers at Traffic

Intersections, Journal of Assistive Technologies.

Barlow J et al., Tabor, 2003, Accessible pedestrian

signals: Synthesis and guide to best practice, National

Cooperative Highway Research Program.

Brabyn J et al., 1933, Talking signs: a remote signage,

solution for the blind, visually impaired and reading

disabled, Engineering in Medicine and Biology

Society.

Dragan et al., 2011, Zebralocalizer: identification and

localization of pedestrian crossings, ACM MobileHCI.

Ophir et al., 2009, Cognitive control in media

multitaskers, Proc. of the National Academy of

Sciences.

Sivaraman, Trivedi, 2010, A general active-learning

framework for on-road vehicle recognition and

tracking, IEEE Transactions on Intelligent

Transportation Systems.

Zoltan et al., 1996, A hierarchical Markov random field

model and multi-temperature annealing for parallel

image classification, GRAPHICAL MODELS AND

IMAGE PROCESSING.

Crammer et al., 2010, On the Algorithmic Implementation

of Multiclass Kernel-based Vector Machines, J. of

Machine Learning Research.

Outdoor Context Awareness Device That Enables Mobile Phone Users to Walk Safely through Urban Intersections

533