Vision Substitution with Object Detection and Vibrotactile Stimulus

Ricardo Ribani and Mauricio Marengoni

Universidade Presbiteriana Mackenzie, S˜ao Paulo, Brazil

Keywords:

Vibrotactile, Vision Substitution, Deep Learning, Object Detecti on.

Abstract:

The present work proposes the creation of a system that implements sensory substitution of vision through a

wearable item with vibration motors positioned on the back of the user. In addition t o the developed hardware,

the proposal consists in the construction of a system that uses deep learning techniques to detect and classify

objects in controlled environments. The hardware comprise of a simple HD camera, a pair of Arduinos, 9

cylindrical DC motors and a Raspberry Pi (responsible for the image processing and to translate the signal

to the Arduinos). In the ﬁrst trial of image classiﬁcation and localization, the ResNet-50 model pre-trained

with the ImageNet database was tried. Then we implemented a Single Shot Detector with a MobileNetV2

to perform real-time detection on the Raspberry Pi, sending the detected object class and location as deﬁned

patterns to the motors.

1 INTRODUCTION

The World Health Organization estimates 285 million

impaired people in the world, being 246 million with

low vision and 39 million completely blind (World

Health Organization, 2012). These people can incre-

ase their environment perception by using technolo-

gies that converts visu al information to different sen-

sors like hear or touch.

The brain plasticity (Bach-y-Rita and Kercel,

2003) has shown the capability of the brain to adapt

to different pa tterns, no matter its source. With some

limitation, the human brain can replace a lost sensor

by using information from other existing sensors in

the body. Some researchers discuss about this phe-

nomena and the sensory substitution (Novich, 2015;

Bach-y-Rita a nd Kercel, 2003; Visell, 2009) and they

cite the ca pability of the brain to change his organi-

zational structure to recognize different patterns. A

common example of this structura l change is the easi-

ness of visually impaired people with the braille wri-

ting system.

Novich ( 2015) has presented a method to allow

deeply dea f users to recognize words from a limited

vocabulary, using encoded audio informatio n captu-

red with a cell phone and sending it to the a wearable

vest that activates a series of vibration mo tors distri-

buted along the user’s back.

The present work show s a prototype that uses

some portable hardware in conjunction with a simple

camera and some vibration motors to send encoded

visual information to the users. The inf ormation co-

mes f rom an o bject detection model and is encoded

using a ﬁxed dictionary. The goal o f this device is not

to completely substitute the vision or other assistance

tools (like th e white ca ne). Since the device is wea-

rable, the user can use the vest and the cane in con-

junction a dding more information about the environ-

ment. A limitation of the white cane is the detection

of objects that are at the top of the user’s ﬁeld of view,

like a tree branch th at can be detected by a camera.

Additionally to the developed hardware, the sy-

stem has an image processing module to perform in-

ference of deep learning models to detect and classify

objects in the scene. In our ﬁrst trial to encode visual

informa tion to the vest we performed some tests with

classiﬁcation models that uses Global Averag e Pool-

ing lay e rs, like the ResNet-50 model (He et al., 2015).

The idea behind trying a classiﬁcation model was to

have a lighter model in a portable device like a Rasp-

berry Pi, but the limitations on encod ing the informa-

tion made us to use an object detection model. The

chosen model fo r this task was the Single Shot Detec-

tor (Liu et al., 20 15) with a MobileNet V2 ( Sandler

et al., 2018 ) and the performance of this model on the

Raspberry Pi exceeded our expectations.

The paper is organized as follows: in section 2 we

present the related work; in section 3, we describe the

architecture of our prototype, including th e hardware

and software; ﬁnally in section 4, results and future

work are discussed.

584

Ribani, R. and Marengoni, M.

Vision Substitution with Object Detection and Vibrotactile Stimulus.

DOI: 10.5220/0007577205840590

In Proceedings of the 14th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2019), pages 584-590

ISBN: 978-989-758-354-4

2 RELATED WORK

In the ﬁeld of sensory substitution just a few use

image captu re and tactile stimulus (Cancar et al.,

2013; Dakopoulos and Bourbakis, 2008; Pereira,

2006; Johnson and Higgins, 2006; Bach-y-Rita et al.,

1969). The existing tr ials to enco de image to touch

consists in simple methods, like downscale the image

to the size of a matrix of a ctuators o r downscale the

edge image. In some other works that has a focus

on navigation (Cancar et al., 2013; Dakopoulos and

Bourbakis, 200 8; Cardin et al., 2007; Johnson and

Higgins, 2006; Meers and Wa rd, 2005) the signal is

a basic function that tells the user the distance from

objects in the scene and they norma lly use depth sen-

sors. In this cases each motor is activated according

to a point of depth giving the user a spatial perception

of the environment to avoid o bstacles. Considering

the brain capability of adapt, these methods are very

modest ones.

Some other projects also uses image capture but to

generate a different type of output, like audio (Saina-

rayanan et al., 2007; Hub et al., 2004; Gonz´alez-Mora

et al., 1999; Meijer, 1992) or touch using gloves ( Lin

et al., 2012; Meers and Ward, 2005) sending signals

to the the ﬁngertips. All these works that uses diffe-

rent sensors to encode image are evidence that doesn’t

matter the origin of the sig nal, the brain learns how to

deal with the information.

As the goal of this work is to evaluate the brain

capability to classify detected objects using touch, be-

low we describ e the most relevant projects to our re-

search.

2.1 Tactile Vision Substitution System

(TVSS)

Deve loped by Bach-y-Rita et al. (1969), this work is

an early example in the ﬁeld o f tactile sensory sub-

stitution. Mounted in a dentist chair, it has a 20 x20

vibration motor matrix that projec t the ima ge captu-

red by an analogical TV camera. No preprocessing is

performed on the image.

The TVSS experiments was pe rformed with 6

users, a totally blind person since 4 years old and ot-

her 5 person totally blind since birth. The ﬁrst ex-

periments included identiﬁcation of vertical lines, ho-

rizontal lines, diagonal lines and curves. Af te r 20h-

40h of training, the users started to recognize the line

orientation with 100% of accuracy. A next experi-

ment comprised recognition of comb ined lines and

objects from 25 classes, like telephone, chair and cup.

The authors didn’t published the detailed accuracy for

object recognition but exp lained that the users were

Figure 1: Tactile Vision Substitution System (Bach-y-Rita

et al., 1969).

capable of recognize perspective variation and dis-

tance variation in function of the size of the ob je c t.

2.2 PhD Thesis of Mauro Pereira Conti

This research was presented by Pereira (2006) from

S˜ao Paulo University and resulted in a prototype that

performs vision substitution using a camera and an

electro-tactile matrix positioned in the user’s abdo-

men. The image is captu red using a simple camer a

and a customized hardware was developed to activate

the electrodes. T he input image is processed using a

PC for edge detec tion, the edges image is downsca-

led to the size of the matrix an d send to the electrodes

matrix.

Figure 2: System developed by Pereira (2006). Top: Con-

trolled board that activates the electrodes. Middle: Main

board with controller boards connected. Bottom: Electro-

des matrix worn by the user.

The author cites the limitations of real-time pro-

cessing and the lim itation regarding to the hardware

Vision Substitution with Object Detection and Vibrotactile Stimulus

585

size. In the Figure 2 is possible to see big control-

ler boards that makes the system unfeasib le to use in

everyday activities. T he use electrodes also requires

the use of a conductive gel to improve conductivity.

In the ﬁr st experiment, users were subjected to r e -

cognize lines. The group of users with people blind

since birth scored 88% of right answers and the group

with normal vision people score d 70%. In the se-

cond experiment the users were presented with com-

plex geometric symbols, like letter L, square, triangle

and circle. T he group of users with people blind since

birth scored 80% of right answers and the group w ith

normal vision people scored only 44%. The last ex-

periment was intended to show different objects to the

users and all the users scored less than 30% correct.

2.3 BrainPort

The BrainPort V1 00 Vision Aid (Stronks et al., 2016)

is a commer cially available d evice that captures the

image using a simple HD came ra and outputs the sig-

nal of the downscaled image into a 20x20 electro-

tactile ton gue display. According to th e product spe-

ciﬁcation, after some training the user is capable of

identify light variation, detect simple objects, recog-

nize small words and detect movements. This system

is an evolution o f the TDU (Tongue Display Unit),

developed by Bach-y-Rita et al. (1998). The price of

this system is around US$10,000.00 and is available

at USA, Europe and Hong Kong.

The system compr ises a portable processing unit

(Figure 3) that receives images from a came ra moun-

ted in a pair of glasses a nd allows the user to adjust

zoom and contrast. Th e system converts the image to

grayscale and reduces it size to 20x20, mapping each

pixel to a point in the electrodes matrix. The brighter

values are responsible to ac tivate the electrodes in a

higher voltage and the darker values will generate lo-

wer voltage. The Figure 4 shows the captured image

and the output image.

Different kinds of tests were made to evaluate the

BrainPort system, including object identiﬁcation, text

recogn ition, light variation, contrast and mobility. We

highlight here the results obtained for object identi-

ﬁcation and text recogn ition. The experiments we re

made with 18 visually impaired people and the goal

was to identif y if the can identify obje c ts in 4 classes:

ball, banana, text marker and cup. After 15–20 hours

of ge neral device training, the subjects had an average

correct rate of 75% in the object recognition task. For

the word recognition task, 10 words were used having

3-5 letters and the subjects had an average correct rate

of 1 5%. The author relates the poor results to the low

resolution ima ges.

Figure 3: (a) The BrainPort V 100 system. (b) B rainPort in

use by a user. (Str onks et al., 2016).

Figure 4: Example of image captured using the BrainP ort

system and the 20x20 output image. (Stronks et al., 2016).

In addition to the research experiments, clin ical

trials were conducted to allow commercial use appro-

val. The same tasks were applied to 75 totally b lind

individuals. After 10 hours of trainin g, the average

correct rate for object identiﬁcation was 91% and for

word recognition was 58%.

2.4 VEST

Our prototype is very similar to the VEST (Figure

5). Although it’s a system developed for hearing-

impaired a nd not visually impaired, the work of N o-

vich (2015) stands out by using a more elaborate met-

hod to encode the audio information in a vest with 26

vibro- tactile actuators and expla ins that the inf orma-

tion encoded in these actuators must b e the result of

a function that the brain is capable of decoding wit-

hout generating confusion about the input signal. The

results obtained with the VEST proved the brain ’s abi-

lity to understand com plex patterns.

To e ncode the information from sound to touch

the author uses the k-means clustering method trai-

ned with english sentences. The number of centroids

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

586

Figure 5: VEST System Novich (2015).

is equal to the number of vibration motors in the vest

and the algorithm is a pplied in certain intervals of

the in put audio data , resulting in 26 frequ e ncy bands

mapped to the moto rs.

After 12 days of training with a 48 words dicti-

onary and after achieving a corre ct rate of 75% in

the tra ining set, the subjects were submitted to a test

set with different words. The performance ranged ap-

proxim ately 35% to 65%, considered a very good re-

sult since the individuals haven’t never seen th e test

set patterns before.

3 PROTOTYPE

Considering the tasks of identify objects in a unknown

environment without any assistance and based on the

theory of the brain’s plasticity, our work proposes a

system similar to the one developed by Novich (2015)

but focused on visually impaired people. The system

proposes vision substitution using a vest with vibra-

tion motors positioned in the back. One of the reasons

for choosing to work with a vest was due to the possi-

bility of the user performing day-to-day tasks without

obstruction of a ny other sense.

Tapu et al. (2014) describes some requirements

needed for an electronic travel aid. T hese require-

ments can be applied to our solution and they are pre-

sented in the Table 1.

3.1 Hardware

In the development of the ﬁrst prototype, we tried to

meet the maximum possible of requirements quoted

by Tapu et al. (2014). The system created is weara-

ble, portable, reliable, inexpensive, u ser friendly and

does not use cables. The robustness will not be evalu-

ated at ﬁrst, because the tests will be done in control-

led env ironments. Since it is a prototype, unexpected

situations can oc c ur even in these e nvironments and

user safety will be prioritized.

The ﬁrst prototype consists of a vest w ith 10 cylin-

drical vibration motors positioned in the back, a pair

of Arduinos model Uno R3, a Raspberry Pi 3 model

B+ and a simple HD camera.

Table 1: Requirements needed for an electronic travel aid.

(Tapu et al., 2014).

Requirement Description

Real-time The system should promptly

send me ssag e s to the user as

soon a s they’re processed.

Wearable It should be worn by the

user. The ears and hands

should be free.

Portable It sho uld be lightweight and

easy to mount, which can

be carried over long d is-

tances, sma ll and ergonomi-

cally shaped.

Reliable Must have a go od cor-

rect rate and reca ll evalua-

tion. However, it must also

have corr ection functions for

unexpected situations.

Low cost It should be comme rcially

accessible to users.

Friendly Simple to use, easy to learn,

no long and exp e nsive wor-

kouts.

Robust The device must resist to d if-

ﬁcult environments and mi-

suse.

No cab les TThere shouldn’t be wires

that limit the user’s mobility.

The Raspberry Pi board was chosen b ecause it is

a low-cost processing unit and porta ble enough to be

carried by the user. This card has a ARM Cortex-

A53 64-bit processor w ith 4 cores of 1.2 GHz, 1 Gb

of RAM and is capable of running different opera-

ting systems. The Raspbian distribution of the Linux

operating system was speciﬁcally developed for this

hardware. The Raspbian system is a lightweight and

capable of r unning the TensorFlow software, which

was used for the infe rence proce ss of the deep lear-

ning models.

The choice of the Arduino boards was mainly due

to the ease of prototyping and the nu mber of PWM

ports available to activate the motors. The Raspberry

Pi card has on ly 4 PWM ports, which is little for th e

proposed design. Ther e fore, the task of activate the

motors was cente red on the Arduino board s and the

inference of the models on th e Raspberry Pi.

The cylindrical motors were used due to the fast

response between fully active a nd fully static, also it’s

easy to control the frequency using a standard voltage

of 3V to 5V in the model used. Each motor can be

activated or deactivated to obtain the desire d n umber

of actuators in operation according to the experiment.

Vision Substitution with Object Detection and Vibrotactile Stimulus

587

In addition, it is also possible to change the positions

of the motors, since they are ﬁxed with velcro strips,

which makes it possible to carry out tests with diffe-

rent conﬁgurations, a s can be observed in Figure 6.

Figure 6: Front facing image of the prototype showing the

internal part. In this image, the motors are conﬁgured as a

3x3 array.

The motors were mounted in the intern al back of a

cycling vest. T he vest is made of a an elastic material,

so that they stay close to the body and give the user

the sen sation of vibrations and differentiation whe n

the motor is on or off . For powering the motors an

auxiliary battery was necessary, since the current su p-

plied b y the Arduino was n ot enough. The Figure 7

shows how the vest is when it is worn. The Arduino

boards, th e battery and the Raspberry Pi are placed in

the back pockets. The camera is attached to a regular

eyew ear and connected to the Raspberry Pi.

The camera is connected to the Raspberry Pi

through USB. The captured image is processed using

Te nsorﬂow to perform inference of a deep learning

model. The output of th e model is post-processed and

sent to the Ard uinos. The signal sent indicates which

motors should be activated. E ach Arduino board has

only 5 PWM ports in the Uno R3 model, so we nee-

ded two boards connected through the I2C protocol.

The ﬁgure 8 presen ts a diagram of the prototype that

was developed.

3.2 Software

3.2.1 Classiﬁcation a nd Localization using

ResNet-50

In our ﬁrst attempt to encode image to touch we tried

the ResNet-50 (He et al., 2015) m odel pre-trained

with Image N et. Our initial idea was to use a ligh t-

weight classiﬁcation model that can also output the

spatial location of the object. The ResNet-50 model

uses global average pooling layers allowing us to es-

timate the position of th e classiﬁed object using the

process described by Zhou et al. (2015) . Each activa-

tion map of the layer before the global average pool-

ing works as a pattern detector in the image and the

(a) Front

(b) Back (c) Camera detail

Figure 7: Our vest prototype. The boards and batteries are

stored inside t he back pockets of t he vest.

Figure 8: System architecture.

weights that connect the last two layers of the model

represent the contribution of each of these patter ns. In

order to obtain th e activation map that contains the lo-

cation of the classiﬁed ob ject, we sum the outpu ts in

the activation layer weighte d with the contributions in

the last layer.

For testing purpo ses we selected two objects and

deﬁned a ﬁxed signal that activates each object. The

ﬁrst object was a guitar, that activates the vibration

motors horizon ta lly from left to right. The second ob-

ject was a laptop, that activates from right to left. We

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

588

got a pre-train e d model with the ImageNet dataset.

The output of the global average pooling layer in

the ResNet-50 m odel has a 7x7 shape. As we have a

limit of 10 motors in our prototype we needed to re-

duce the size of the global average pooling to 3x3 to

send the informatio n regarding the location of the ob-

ject. A new route was created after the existing global

average p ooling layer with one more global average

pooling layer, downsizing the activation map to 3x3.

The vibration motors in the vest were conﬁgured as

the same, allowing us to map th e signal directly to it.

3.2.2 Object Detection with SSD

Even estimating the location of the object through

global ave rage pooling layers, the use of a classiﬁ-

cation model has limitation s fo r the proposed appli-

cation. There is no information regarding the num-

ber of detected objects and it’s always predicting the

class with the greatest pro bability. This is critical in a

system that sends the signal to vibr ation motors, be-

cause the signals are sent to the motor s all the time

causing confusion to the user. To work around this

problem, w e started using an object detection model.

The chosen model was the Single Shot Detector (Liu

et al., 2015) with MobileNetV2 (Sandler et al., 2018)

as bac ken d that has a good balance between perfor-

mance a nd accuracy, 22 mAP in the Microsoft COCO

dataset (Lin et al., 2014) and average 1 FPS running

inference on the Raspberry Pi.

The Microso ft COCO database has 90 classes.

However, in order to attend this experiment we used

objects of only 4 classes (cu p, remote control, scissors

and bottle) and the layout of the motors was modiﬁed

in relation to the previous experiment, from 3x3 rows

and columns to 4x2. The activation of the vibration

motors is done per row, where each pair of motors in

each row is activated when detecting an object among

these 4 classes.

Since th e setup of the motors was modiﬁed , we

also changed the output signal indicating the position

of the object. To map the location with r e spect to the

input image, it was divided into 8 quadrants according

to the conﬁguration of the vest, in 4 rows and 2 co-

lumns. Since the m odel already provides the position

and size of the bounding box, we calculated the area

of it in each quadrant obtaining th e intensity value for

each m otor, as can be seen in Figure 9.

Figure 9: Detected object (cup) and the values of activation

for the vibration motors for class and position. Left: Input

image with the 4x2 grid and the detected object. Center:

Values send to the vibration motors to indicate the cup class.

This value is hardcoded fr om a deﬁned dictionary. Right:

Values send to the vibration motors to indicate the position

of the object.

4 DISCUSSION AND FUTURE

WORK

In the ﬁrst attempt w e used a classiﬁcation model in

which it was possible to perfor m a simple classiﬁca-

tion in the image and take advantage of the globa l

average pooling layers to locate the object. However,

this model has a limitation for the proposed use case,

mainly by constantly activating the vibration moto rs

that can confuse the user. In o rder to solve this pro-

blem, we decide d to use the SSD obje c t detection mo-

del.

From the tests performed in our labor atory using

the vest with the SSD object detection model, it was

possible to send signals to the vest that repr esents the

objects and the position of this o bjects in sequence.

By using the vest with a real-time detec tion, it was

possible to clearly perceive the difference between

signals sent that r epresents d ifferent objects. The sig-

nal regarding the location of the object seemed a little

more con fusing at ﬁrst, but after a few minutes of tr ai-

ning the signal started to make more sense. This phase

was intended to check the system operation, including

the communication between all the parts and valida-

tion o f the signal send to the vest, that will enable the

next steps of our research. No evaluation were made

with visually impaired individuals yet.

Novich (2015) explains that the encoded signal

must be a function for the brain to be able to lea rn to

decode the input. Considering this, adjustme nts can

be made in the object detection model to create out-

puts that are functions of th e learned features in the

hidden layers of these models. That is, a possible op-

tion for a next version is to use the f eatures learned by

the network to generate a pattern that is similar when

the input images are similar. To achieve this goal we

Vision Substitution with Object Detection and Vibrotactile Stimulus

589

will perfo rm experiments with clustering algorithms

(such a s k-means), dimensiona lity reduction layers to

match the number of vibration motors, among o ther

techniques that will be investigated .

After having consistent results using the vest in

the laboratory, we plan to perform experiments with

visually impaired users.

ACKNOWLEDGEMENTS

The authors thank Fundo Mackenzie de Pesquisa

(Mack-p esquisa) fr om the Universidade Pre sbiteriana

Mackenzie for the ﬁnancial support for this research.

REFERENCES

Bach-y-Rita, P., Collins, C. C., Saunders, F. A., White, B.,

and Scadden, L. (1969). Vision substitution by tactile

image projection. 221:963–4.

Bach-y-Rita, P., Kaczmarek, K. A., Tyler, M. E., and

Garcia-Lara, J. (1998). Form perception with a 49-

point electrotactil e stimulus array on the tongue: A

technical note. Journal of Rehabilitation Research

and Developm ent, 35(4):427–430.

Bach-y-Rita, P. and Kercel, S. W. (2003). Sensory substi-

tution and the human–machine interface. Trends in

Cognitive Sciences, 7(12):541 – 546.

Cancar, L., Diaz, A., Barr ientos, A. , Travieso, D., and Ja-

cobs, D. M. (2013). Tactile-Sight: A sensory substitu-

tion device based on distance-related vibrotactile ﬂow

regular paper. International Journal of Advanced Ro-

botic Systems.

Cardin, S., Thalmann, D., and Vexo, F. (2007). A wearable

system for mobility improvement of visually impaired

people. Visual Computer.

Dakopoulos, D. and Bourbakis, N. (2008). Preserving vi-

sual information in low resolution images during navi-

gation of visually impaired. In Proceedings of the 1st

International Conference on PErvasive Technologies

Related to Assistive Environments, PETRA ’08, pages

27:1–27:6, New York, NY, USA. ACM.

Gonz´alez-Mora, J., Hern´andez, A. R., Ramos, L. F. R.,

Dfaz-Saco, L., and Sosa, N. (1999). Development of

a new space perception system for blind people, based

on the creation of a virtual acoustic space.

He, K., Zhang, X., Ren, S. , and Sun, J. (2015). Deep

residual learning f or image recognition. CoRR,

abs/1512.03385.

Hub, A., Diepstraten, J., and Ertl, T. (2004). Design and de-

velopment of an indoor navigation and object identiﬁ-

cation system for the blind. In Proceedings of the 6th

International ACM SIGAC CESS Conference on C om-

puters and Accessibility, Assets ’04, pages 147–152,

New York, NY, USA. ACM.

Johnson, L. A. and Higgins, C. M. (2006). A navigation

aid for the blind using tactile-visual sensory substitu-

tion. In 2006 International C onference of the IEEE

Engineering in Medicine and B iology Society, pages

6289–6292.

Lin, K. W., Lau, T. K., Cheuk, C. M., and Liu, Y. (2012). A

wearable stereo vision system for visually impaired.

In 2012 IEEE International Conference on Mechatro-

nics and Automation, pages 1423–1428.

Lin, T., Maire, M., Belongie, S. J., Bourdev, L. D., Girshick,

R. B., Hays, J., Perona, P., Ramanan, D., Doll´ar, P.,

and Zitnick, C. L. (2014). Microsoft COCO: common

objects in context. CoRR, abs/1405.0312.

Liu, W., Anguelov, D., Er han, D., Szegedy, C., Reed, S. E.,

Fu, C., and Berg, A. C. (2015). SSD: single shot mul-

tibox detector. CoRR, abs/1512.02325.

Meers, S. and Ward, K. (2005). A substitute vision system

for providing 3 d perception and gps navigation via

electro-tactile stimulation.

Meijer, P. B. L. (1992). An experimental system for au-

ditory image representations. IEEE Transactions on

Biomedical Engineering, 39(2):112–121.

Novich, S. D. (2015). Sound-to-Touch Sensory Substitution

and Beyond. Master’s thesis, Rice Universit y.

Pereira, M. C. (2006). Sistema de substituic¸ ˜ao sensorial

para aux´ılio a deﬁcientes visuais via t´ecnicas de pro-

cessamento de imagens e estimulac¸ ˜ao cutˆanea. PhD

thesis.

Sainarayanan, G., Nagarajan, R., and Yaacob, S. (2007).

Fuzzy image processing scheme for autonomous navi-

gation of human blind. Appl. Soft Comput., 7(1):257–

264.

Sandler, M., Howard, A. G., Zhu, M., Zhmoginov, A., and

Chen, L. (2018). Mobilenetv2: Inverted residuals and

linear bottlenecks. CoRR, abs/1801.04381.

Stronks, H. C., Mitchell, E. B., Nau, A . C., and Barnes, N.

(2016). Visual task performance in the blind with the

BrainPort V 100 Vision Aid.

Tapu, R., Mocanu, B., and Tapu, E. (2014). A survey on we-

arable devices used to assist the visual impaired user

navigation in outdoor environments. In 2014 11th In-

ternational Symposium on Electronics and Telecom-

munications (ISETC), pages 1–4.

Visell, Y. (2009). Tactile sensory substitution: Models for

enaction in hci. Interact. Comput., 21(1-2):38–53.

World Health Organization (2012).

Zhou, B., Khosla, A., Lapedriza,

A., Oliva, A., and

Torralba, A. (2015). Learning deep features for dis-

criminative localization. CoRR, abs/1512.04150.

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

590