obstacledetectioninimagesliteraturesurvey-140317210549-phpapp01

Obstacle Detection in Images

T.H .H Methmal (Reg No. 2010/CS/181, Index No. 10001816)

University of Colombo School of Computing

[email protected], 0711361409

A Literature Survey

Presented to University of Colombo School of Computing

in partial fulllment of the requirements

for the Degree of Bachelor in Computer Science

SCS3013 Literature Survey Courses

Adviser: Prof. N. D. Kodikara

Used harvard referencing style

4300 of word count

Lyx tool used for format

December 11, 2013

Copyright 2013 by Hasanga Methmal. All rights reserved.

Abstract

In this survey, we research about ways of obstacle detection technics in detail. Study will

go through the importance of obstacle detection, various applications which are included

obstacle detection kits and various technics to achieve obstacle detection. Mainly survey

will discuss about three main areas of obstacle detection. Applications which are innovate

to assist visually impaired people, Path planning using obstacle avoidance of mobile robots

and various kinds of autonomous vehicles has built in obstacle detection system. As well

as survey will review main obstacle detection technics such as monocular technics, stereo

technics, depth maps generating technics etc. In the discussion all the technics will be

compared and consider their pros and cons briey. Later part of the survey discusses

about future enhancements and challenges in obstacle detection.

i

Acknowledgements

I would like to thank my advisor, Prof. N. D. Kodikara , for his experienced guidance

and constant support throughout the period . I also be obligated to thanks to my family

members and friends, whose support, well wishes and encouragements to brought me

here.

ii

Contents

Abstract i

Acknowledgements ii

List of Figures v

1 Introduction 1

1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.2 Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

2 Background 2

2.1 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

2.2 Errors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

2.2.1 False Positive . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

2.2.2 False Negative . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

2.2.3 Noise and Non-Linearities . . . . . . . . . . . . . . . . . . . . . . 3

2.3 3D versus 2D . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

3 Approaches to obstacle detection 4

3.1 Obstacle detection using 3D grid and 2D grids. . . . . . . . . . . . . . . 4

3.1.1 Method introduction . . . . . . . . . . . . . . . . . . . . . . . . . 4

3.1.2 Stereo Measurement . . . . . . . . . . . . . . . . . . . . . . . . . 5

3.2 Obstacle detection by using sequence of lters . . . . . . . . . . . . . . . 7

3.2.1 Ground/Non-ground Segmentation. . . . . . . . . . . . . . . . . . 7

3.2.2 Background Subtraction. . . . . . . . . . . . . . . . . . . . . . . . 7

3.2.3 Size lter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

3.2.4 Appearance-based Recognition Filter. . . . . . . . . . . . . . . . . 8

3.2.5 Shape-based Recognition Filter. . . . . . . . . . . . . . . . . . . . 8

3.3 0bject collision detection algorithm based on disparity images. . . . . . . 8

3.4 Haar-like features with cascade classier . . . . . . . . . . . . . . . . . . 9

3.5 Fast Human Detection for Indoor Mobile Robots Using Depth Images . . 10

3.5.1 Depth image segmentation. . . . . . . . . . . . . . . . . . . . . . 10

3.5.2 Region Filtering and Merging . . . . . . . . . . . . . . . . . . . . 11

3.5.3 Candidate Classication . . . . . . . . . . . . . . . . . . . . . . . 11

3.6 Obstacle detection using Equivalent points. . . . . . . . . . . . . . . . . 11

3.7 Multi-cue Visual Obstacle Detection. . . . . . . . . . . . . . . . . . . . . 12

3.7.1 Geometric obstacle detection . . . . . . . . . . . . . . . . . . . . . 12

3.7.2 Color base obstacle detection . . . . . . . . . . . . . . . . . . . . 12

iii

4 Discussion 13

5 Conclusion 14

6 Future works 15

References 16

iv

List of Figures

3.1 Gate, obstacle, oor and bump regions that robots can move through. . . 4

3.2 Flowchart of the motion planning method . . . . . . . . . . . . . . . . . 5

3.3 Robot can see obstacles . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

3.4 Takes all measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

3.5 3D grid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

3.6 C-Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

3.7 Implemented series of lters, . . . . . . . . . . . . . . . . . . . . . . . . . 7

3.8 Depth image . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

3.9 Algorithm owchart . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

3.10 Identifying the corresponding regions of distinct object in depth images . 10

3.11 Image after omit unnecessary lines . . . . . . . . . . . . . . . . . . . . . 11

3.12 False negative and false positive . . . . . . . . . . . . . . . . . . . . . . . 12

v

1 Introduction

This survey refers how obstacle detection is done by various application areas. Also it

is discussed about technics which are used in obstacle detection in images. Recently ob-

stacle detection is a quit interesting topic among the researchers. Numerous applications

are innovated dierent areas which are included obstacle detection kits. So survey will

discussed mainly about three areas. Most of the vision systems of mobile robots are

contain an obstacle detection system. Recently robots are hugely involving industrial

production and military activities. Navigating through obstacle lled environments is a

big challenge to mobile robots. So there are so many technics to apply to mobile robots

to make easier the navigating task. As well as there are so many applications which are

using obstacle detection system to assist visually impaired people. Those people suering

from lack of information of their surroundings. So assisting applications detect various

obstacles and analyze and present with details to the user. Now days there are lots of

unmanned vehicles are produced for dierent kind of purpose. Such as military, farming

and whether forecasting elds are using many aircrafts and vehicles which are come up

with obstacle detection systems.

1.1 Motivation

Vision base obstacle detection truly motivated by nature. Most animals gain dierent

kinds of vision systems in nature. As humans we got eyes which providing amazing

images. Eyes are giving us remarkable 3D information to us for obstacle detection.

Flying insects have optimal ows which are giving them absurd navigation and obstacle

detection capabilities.

Actually obstacle detection system is an important sensor in robots and autonomous

vehicle. Because it is a very crucial gadget to have when it in the unknown world. It

may deliver very important information about its surrounding to succeed the navigation.

So vision base obstacle detection directly commit to the safe operation of mobile robots

and autonomous vehicle

1.2 Objectives

Main aspiration of this survey is discuss about pros and cons existing algorithms and

analyze their eciency. Research question of this literature survey is the most suitable

real time technic is for navigating and path planning purposes to mobile robots.

1

2 Background

2.1 Applications

Vision obstacle detection is a very interesting topic in recent past. Because there are so

many applications have vision system these days. Mainly mobile robots have obstacle

detection system[1] [2] [7]. When mobile robot enter to an unknown environment his

vision system is very helpful to do their operations and survive in that environment. As

well as obstacle avoidance is help to planning their path and reach to the destination[1].

Especially military services use these kinds of robots for dangerous operations[2]. As well

as vision system use for create applications to help visually impaired people[3] [4] [7].They

are suering from lack of information of their surroundings. Researchers have invented

applications that can detect obstacles and give them a message where the obstacle is.

Staircases are really challenging to go through for a visually impaired people. So there

are applications for detecting staircase with details[4]. There are so many unmanned

autonomous vehicles are using obstacle detection systems. When there is no one to

control that vehicle has to have a good vision to navigate through the obstacles. In

military services use unmanned aircrafts for spy enemy activities. Those helicopters have

immediate landing emergency in case of dierent situations. Then helicopters have to

nd minimum obstacle area to landing[10]. As well as there are applications that creates

for farming purpose. There is an autonomous vehicle that has a vision system to obstacle

avoidance. It can safely navigate through the fruit trees[11] .

2.2 Errors

When we work with train data there can be some errors obviously. In vision obstacle

detection has mainly two kinds of errors.

2.2.1 False Positive

There can be an error that your obstacle detection system output has an obstacle. But

In the real environment there is nothing. For Example if there is a paper on the oor

system may be detecting it as an obstacle. But actually robot can easily walk on that

paper. So it is false positive behavior.

2.2.2 False Negative

There may be an error that truly environment has an obstacle. But your system not

detects it. It is an false negative error. In those situations your devices can be harmed

because of undetected obstacles.

2

2.2.3 Noise and Non-Linearities

Noise is dened as some random dierences in brightness information of the image. Those

noises are generated in lm grains or electronic noise in your input stereo camera. Re-

searchers use dierent technics to eliminate the Noise and Non-Linearities. Random

Gaussian noise and Kalman Filter are some of them.

2.3 3D versus 2D

Many robots have 2D obstacle detection system. Solving 2D obstacle detection problem

is much easier than solving 3D obstacle detection. 2D detection system always deals with

intensity map at each pixel in the particular picture. It check whether obstacles to be

identied and not identied. Very rst image indicates changes in the scene. But in 3D

obstacle detection system have to picture the whole scene.

3

3 Approaches to obstacle detection

3.1 Obstacle detection using 3D grid and 2D grids.

It is very important to know about its environment for beep robot while it is seeking

path to reach a destination [1].This method will show how robot identies the obstacles

and how plan its path through below mentioned obstacle regions.(Figure 1)

Figure 3.1: Gate, obstacle, oor and bump regions that robots can move through.

Specialty of this method is robot can move through above obstacle area and it can

measure the obstacle. First robot capture images its surrounding and then generates a

3D grid using those images. After 3D grid is converted in to a 2D grid.

3.1.1 Method introduction

There are some assumption in propose method.

1. All the objects in the environment divided into four classes such as oor, obsta-

cle, bump and gate areas. So robot can recognize objects by shape and distance

information.

2. Size and location of the landmarks antecedently. So robot knows its correct coor-

dination.

3. Robot knows his destination in advance. So it plans the path to destination.

Flowchart of the motion planning method is mention below.(Figure 2)

4

Figure 3.2: Flowchart of the motion planning method

3.1.2 Stereo Measurement

Robot gets information about environment using a stereo camera. Figure 3 shows images

of the environment. Then it will convert into the 3D model of the environment as shown

in Figure 4 .Then robot can get all the measurements to build the 3D grid.(Figure 5)

Figure 3.3: Robot can see obstacles

Figure 3.4: Takes all measurements

5

Figure 3.5: 3D grid

Method called conguration space (c space) used for represent the robot. It considers

the robot as a point without any size. As well as all the obstacles are call conguration

obstacles(c-obstacle) in the environment and they are enlarged based on robot size. Figure

6 shows an example of c- space.

Figure 3.6: C-Space

After get all the 3D measurements it will create a 3D grid map as shown in gure

4.Then it converts to a 2D map.2D map can be classied into several region. Figure 6

shows all classied areas in 2D map.

Those unmeasured area will be reduced by continues learning process. Now onwards

2D grid map recognize as a graph which is including nods and graph. Robot can plan its

path by using graph traversing technics.

6

3.2 Obstacle detection by using sequence of lters

This approach is a combination of stereo and monocular image-based vision algorithm[2]

.Mainly this method use a sequence of lters through the color camera lters. First go

through with the image Figure 7.

Figure 3.7: Implemented series of lters,

Then extract information of the foreground grouped into the blobs. Then those blobs

will be ltered based on size model, appearance model and shape model. Several non-

target blobs can be eliminated by this process and system can focus on possible targets.

Most important thing is this method use series of lters and each model that is learn on

the y from the previous model.

3.2.1 Ground/Non-ground Segmentation.

Since this use for mine discovering robot we have to clearly identify the low- lying obstacles

in the surface. For this purpose we use a correlation-based pyramidal stereo algorithm

to get a threshold. We use the RANSAC algorithm to get an estimate ground surface.

Give a series of 3D stereo points to algorithm and get an estimate value. According to

that value we decide if it is a potential target or not.

3.2.2 Background Subtraction.

Main distraction of the ground is moving objects. For overcome this problem algorithm

create a composite image over the object moving time. Also create before and after the

object is thrown. Analyze all the images and all the detected changes grouped into the

blobs.

3.2.3 Size lter

In this phase it lters according to the known size of the target object. Distance to the

blob will be estimated using stereo.

7

3.2.4 Appearance-based Recognition Filter.

We use only consider about the color. Because target objects are texture less. Robot

doesn't' t know how is the target obstacle looks like in advance. So we have to present

all sides of the object to the robot and train to recognize them.

3.2.5 Shape-based Recognition Filter.

Most eective way to identify objects is its shape. Shape base recognition algorithm use

to identify the object from the object classes.

3.3 0bject collision detection algorithm based on disparity im-

ages.

This technics mainly use for detecting obstacles with their depths[3]. This approach was

used for an application which helps for the visually impaired people[1]. They really want

detailed information about their surroundings. We use disparity image or depth map to

storing the depth of each pixel in the image. Each pixel of the depth map corresponding

to a pixel in the image. Near objects can be seen lighter color and far objects are

darker.Figure 8 shows an example.

Figure 3.8: Depth image

Information about depth maps are stored using 64 gray levels. But there are more

levels have to be added in dierent circumstances such as outdoor scenes or high quality

pictures. Depth of a pixel is calculated by following formula.

z =f bd

z is the depth, f is the focal lengths in pixels, B is the base line in meters and d is the

disparity. Proposed algorithm is detected two levels of objects. One meter depth objects

catch as a near closeness and two meter threshold for rst detection.

Proposed algorithm briey

8

1. Breakdown disparity image with 2D Ensemble Empirical Mode Decomposition.

2. Output disparity images ltered to dispose of higher frequency and noise.

3. Dene two regions of interests (ROIs) to detect nearby objects in two levels. in this

case we got one and two meters.

Finally combine all the information of depth maps and apply into real image to excerpt

real object information.

3.4 Haar-like features with cascade classier

When a visually disable person walking through an area staircase is a major obstacle for

him to go through[4]. So there are application which detect various staircases to assist

them.[1]

Main object of the technic is improving the accuracy of detection. To achieve this

target we have to collect set of positive image set which are containing 4 to 5 steps and

photos should be taken on various conditions. As well as we have to collect set of negative

samples also which not contain any staircase. After that training is done until calcied

into 18 layers. Note that training is performed in fewer steps than a normal application.

Because always good to have a higher detection rate in staircase detection applications.

When detection staircase most of time detect similar regions like staircase and they call

candidate detection. For overcome this false detection, we use cascade classiers create

from training session and the 40 * 40 sub windows with dierent sizes for scanning. Haar

detector creates multiple hits in the staircase region and false detection has single isolated

detection.

Note that Stereo camera system produces very poor results for staircase detection

although it gives very informative results with less false detection rate. That's why we

have to come up with such algorithm.All the stages of the algorithm is shown in Figure

9.

Figure 3.9: Algorithm owchart

9

3.5 Fast Human Detection for Indoor Mobile Robots Using Depth

Images

One of the biggest challenges for mobile robots is detect moving humans[5]. Because

Humans can be appear in dierent postures like walking, sitting. Even some are seen in

partially. So detecting those various kinds of humans is very challenging task.

For this approach we use Microsoft depth camera to detect humans. So depth camera

is gives depth images. This algorithm is scanning the depth images and identies the

human in it. Mainly algorithm has three phases.

3.5.1 Depth image segmentation.

Figure 3.10: Identifying the corresponding regions of distinct object in depth images

Purpose of this step is identifying the corresponding regions of distinct object in depth

images.(Figure 10) But resolution of depth images get poor with the distance. So that

is not necessary to segment those images at larger distance. Accurate way is sub sample

the depth image and then does the segmentation.

10

3.5.2 Region Filtering and Merging

Figure 3.11: Image after omit unnecessary lines

In this phase algorithm take Figure 10 as input and omit unnecessary lines which are

highly unlikely humans and merge into the remaining regions.(Figure 10)

3.5.3 Candidate Classication

In the nal step we use a support vector machine to decide whether the detected object

human or not.

3.6 Obstacle detection using Equivalent points.

Firstly we have to take a photo of an object by using two cameras this object will specify

a particular place for itself in the pictures taken by cameras[6]. We named these points as

equivalent points. If someone takes picture from dierent distance those places no longer

equivalent.

Then we have to convert the image into binary mode by changing value of pixels into

0 and 1. After that we separate equivalent region from others and blacken the rest of

image. We should identify the object which we interested about and draw a box around

them. Since we take those two images using two cameras there are regions that are exist

in rst picture and not exist in second one. We have to cut those regions out. Since we

are using two cameras in two positions to take the photo one might be bigger than other.

So we have to stretch the images and make them standardize to comparable. Researcher

has found three regions that object can exist. In this approach we can tell object whether

exist or not and the region that obstacle exist.

11

3.7 Multi-cue Visual Obstacle Detection.

Visual detection algorithms are mainly base on color and geometric information[7]. Color

base only approaches are copiously gives false negatives and false positives. If there is

a paper of another color on the oor robot will detect it as an obstacle. So it is a false

positive. As well as if there is obstacle which is same color as oor robot will not detect

it. It is false negative. Example instances have been mentioned in below Figure 12.

Figure 3.12: False negative and false positive

Even geometric only approaches can be given many false negative results. This ap-

proach has two kinds of obstacle detectors both color base and geometric base. There are

dierent approaches to detect them. Both algorithms will produce binary images that

are consisting of white areas which represent the obstacles.

3.7.1 Geometric obstacle detection

For do this we have to make an assumption that oor is nearly planar. In the algorithm

two camera images are map in to another by 3d points. Algorithm determines a new

position for each pixel point. Obstacle detection is done wrapping one of the images to

other and comparing two images. But higher lightning conditions algorithm may give

more false positive results.

3.7.2 Color base obstacle detection

First of all we have to take several images of the ground and do the training process.

Training is done selecting three region of the ground and all the result put into a three

dimensional histogram. Classication process divides into two phases. In rst stage check

every pixel with particular value in histogram. If value is less than the threshold detected

as obstacle. After that those pixels are check again as a 4 * 4 pixel block. If those all

pixels are checked as an obstacle then determine as an obstacle.

12

4 Discussion

As I mentioned in chapter 03 there are many approaches to detect obstacle detection

in images. These days there are lots of sensors are using in commercial devices like

unmanned military aircrafts. But those sensors are very expensive. Because of that many

devices have a pair of stereo cameras and most of the time running an ecient algorithm

on stereo images. When I do the literature survey I identied few main obstacle detection

algorithms that often use in applications.

Obstacle detection using depth maps and disparity image is quit famous method

among the researchers. When we got a Disparity images it should be ltered by a proper

algorithm to reduce the noise. In [3] is use 2D EEMD algorithm to lter the noise.

Problem is this method after two layers again it similar to rst layer. Then it is very

dicult to user to identify the place of the object. Other than that user can gain more

details from the depth map. User can determine the shape of the obstacle as well as

the depth of the obstacle. But in high lightning area is not given very accurate results.

Human detection application in [5] also uses depth image segmentation. Since it detect

only human obstacles task is much harder than other approaches. But in the end both

two applications success and obstacle detection is done with minimum error percentage.

In mobile robot path planning application using 3d and 2D grid maps obstacle de-

tection is done successfully than depth maps[1].Because Robot creates a 3D grid map all

over the environment and build a 2d grid map according to that. In this case robot know

not only his front but also he knows what are the objects that around him. Problem is

when not complete the training process there may be some unknown objects around the

robot. Other than that this is very ecient path planning approach to robots.

Object detection using ltering is a not very ecient approach to real time appli-

cations. Because it has many steps and can be given many false negative results. As

well as it can be trained only for very few items at once. This approach use for mine

explore robot in this literature survey[2].In the appearance model obstacle detection done

by color detection. I think it very inecient approach to detect a mine.

Because mines are existing in dierent colors. Shapes of the mines almost same. But

again there can be so many exceptions. When your robot play with mine like dangerous

things it should be 100% present accurate. Otherwise may be destroyed your application.

13

5 Conclusion

As I mention earlier obstacle detection is most important topic in many elds. Among

those elds robotic technology, unmanned aircrafts in military services and Applications

that are helps to blind people are very important. I must say this is a much brought

area in vision and recent past many researchers are involved to this elds to explore new

methods and approaches.

Although have many technics and algorithms to obstacle detection there are some

suitable approaches depends on the application which are use them.

In mobile robot area most of the time use 3D and 2D grid maps [1] and depth images[5].

I think most ecient way of path planning and obstacle avoidance technic is using the

grid maps. Because in grid maps robot train for his current environment and he knows

the all the obstacles. But only disadvantage is when we add a new obstacle in to the

environment robot have to generate all the grid maps again. In depth images robot real

time detect the obstacle and path his plan. But this approach much slower than grid

maps.

There are two approaches mainly discussed in applications about blind people. First

one is using depth images[3]. I think this is a good approach because user can get all

the information that which shape obstacle is and how many meters to the obstacles.

Problems are the accuracy of the algorithm depends on image brightness and pixel size.

Advantage is user can see the obstacle in two depths. When user knows the obstacle

exist in two meters away he has a time to change his path. Other approach is using

Equivalent Points. But this method detects only the obstacle. Depth can't be measured.

So considering these two approaches you can realize using the depth map will be a better

approach.

14

6 Future works

This literature survey has described how obstacle detection done by using a camera.

Besides of that there are so many obstacle detection sensors in the market. But they are

very expensive at the moment. That's why many application running with stereo camera.

For an example LADAR (Laser Detection and Ranging) sensors are using for un-

manned aircrafts for emergency landing. They are scanning the whole area before landing

the aircraft[12]. LADAR system using for determine the distance to an object. Benets

of the LADAR system can get the image of the target while it is calculating the distance

for the target. That functionality can use for obstacle detection purposes.

15

References

[1] Atsushi Yamashita, Masaaki Kitaoka, and Toru Kaneko, 2011. Motion Planning of

Biped Robot Equipped with Stereo Camera Using Grid Map. IEEE Transactions on

Robotics and Automation, Vol.5 No.5, 639-647.

[2] Stancil, Brian A., Hyams, Jerey, Shelley, Jordan; Babu, Kartik; Badino, Hernn.;

Bansal, Aayush; Huber, Daniel; Batavia, Parag, 2013. CANINE: a robotic mine

dog. Intelligent Robots and Computer Vision XXX: Algorithms and Techniques.

Proceedings of the SPIE, Volume 8662, article id. 86620L, 12.

[3] Paulo Costaa, Hugo Fernandesb, Paulo Martinsc, Joo Barrosod, Leontios J. Had-

jileontiadise, (2012). Obstacle detection using stereo imaging to assist the navigation

of visually impaired people. In 4th International Conference on Software Develop-

ment for Enhancing Accessibility and Fighting Info-exclusion. Douro Region, Portu-

gal, July 19-22, 2012. Portugal: Procedia Computer Science. 83-94.

[4] Young Hoon Lee ,Tung-Sing Leung, Grard Medioni, (2012). Real-time staircase

detection from a wearable stereo system. In Pattern Recognition (ICPR), 2012 21st

International Conference. Tsukuba, 11-15 Nov. 2012. 1-4.

[5] Benjamin Choi, C etin Mericli, Joydeep Biswas, and Manuela Veloso, 2013. Fast

Human Detection for Indoor Mobile Robots Using Depth Images. IEEE International

Conference on Robotics and Automation (ICRA)

[6] Nazli Mohajeri, Roozbeh Raste, Sabalan Daneshvar,(2011). An Obstacle Detection

System for Blind People. In Proceedings of the World Congress on Engineering.

London, U.K., July 6 - 8, 2011. WCE.

[7] Luis J. Manso, Pablo Bustos, Pilar Bachiller and Jose Moreno 2010. Multi-cue

Visual Obstacle Detection for Mobile Robots. journal of physical agents, Vol 4, 1 pg

[8] Tomoyuki Mori and Sebastian Scherer, "First Results in Detecting and Avoiding

Frontal Obstacles from a Monocular Camera for Micro Unmanned Aerial Vehicles,"

International Conference on Robotics and Automation, May, 2013

[9] A. Ess, B. Leibe, K. Schindler, L. van Gool1 , (2009). Moving Obstacle Detection

in Highly Dynamic Scenes. In Robotics and Automation, 2009. ICRA '09. IEEE

International Conference. Kobe, 12-17 May 2009. 56 - 63 pg.

[10] Tarek El-Gaaly, Christopher Tomaszewski, Abhinav Valada, Prasanna Velagapudi,

Balajee Kannan, and Paul Scerri, "Visual Obstacle Avoidance for Autonomous Wa-

tercraft using Smartphones," Proceedings of the Autonomous Robots and Multirobot

Systems workshop (ARMS 2013, at AAMAS 2013), May, 2013.

16

[11] Tarek El-Gaaly, Christopher Tomaszewski, Abhinav Valada, Prasanna Velagapudi,

Balajee Kannan and Paul Scerri, (2012). A Practical Obstacle Detection System for

Autonomous Orchard Vehicles. In 2012 IEEE/RSJ International Conference. Vilam-

oura, Algarve, Portugal, October 7-12, 2012.. zzz: zz. 3391- 3399 pg.

[12] Sebastian Scherer, Lyle Chamberlain, Sanjiv Singh, 2012. Autonomous landing at

unprepared sites by a full-scale helicopter. Robotics and Autonomous Systems.

[13] Saeid Fazli, Hajar Mohammadi D., Payman Moallem, 2010. An Advanced Stereo Vi-

sion Based Obstacle Detection with a Robust Shadow Removal Technique. SOURCE

World Academy of Science, Engineering & Technology, 43 vol, 699p.

[14] 4. Heather Jones, Uland Wong, Kevin Peterson, Jason Koenig, Aashish Sheshadri,

and William (Red) L. Whittaker, "Complementary Flyover and Rover Sensing for

Superior Modeling of Planetary Features," Proceedings of the 8th International Con-

ference on Field and Service Robotics, July, 2012.

[15] Kostavelis, I, Nalpantidis, L & Gasteratos, A 2009, 'Real-Time Algorithm for Obsta-

cle Avoidance Using a Stereoscopic Camera'. in Third Panhellenic Scientic Student

Conference on Informatics.

17