The vision-based driver’s assistance
systems have received considerable attention in the past two decades since many
car accidents are as a result of the drivers’ fatigue or lack of awareness. The
detection of a traffic sign on the road is very useful in improving the driving
thereby helping to reduce the accidents that have been rampant. Thus, the major
purpose of the vision-based traffic sign detection systems is to detect the
yellow and the red signs from the vehicles that are ahead of the use of computer
vision technologies and then making a stop of slow down where necessary. In the intelligent vehicles, the driver
assistant system is designed to assist the drivers to perceive the any
dangerous situations early enough to avoid accidents because the system senses
the danger and warns the driver by reading the traffic signs.
In this paper, the researcher aims at
developing an algorithm that can detect traffic signs based on the color and
shape. This algorithm will make use of images taken by a camera that will be
mounted in front of the moving car. There will be a testing of two types of
traffic signs including the yellow warning signs and the red stop signs after
which the results will be summarized. The study concludes that that color-based
detection can be easily illuminated whereas the shape detection is based on the
complexity of the background. It is
anticipated that this research will help in addressing the problem of accidents
on the roads as it also helps to make the best use of technology to improve the
transport sector. The study will be useful not only to the drivers but also to
all the other stakeholders such as the travelers and the vehicle owners.
CHAPTER 1
INTRODUCTION
1.1. Introduction
In the past three decades of so,
autonomous vehicles have been a subject area of intense research. The state-of-the-art research leverages
complex techniques in computer vision for detection of a traffic sign, which
has been an active area attracting the attention of several researchers. Many research works have been conducted
utilizing a front viewing camera whose purpose is for vehicle localization and
navigation, obstacle avoidance as well as environment mapping. The on-road applications of these vision
detection systems have incorporated lane detection, the occupant pose
inference, as well as driver distraction detection. Some researchers have described that it is
imperative to consider not only the vehicle’s surrounding, and the external
environment when designing the vehicle assistance systems but the internal
environment and the driver should also be taken into account (Trivedi, Gandhi,
& McCall, 2007; Tran & Trivedi, 2012).
The focus on other types of information
while designing the sign detector can make the whole system even better (Morris
& Trivedi, 2010). Every traffic
detection system is aimed at achieving four main goals including. The first one
is that the algorithm should be accurate.
Morris and Trivedi have said that accuracy is the most basic requirement
and the most important evaluation metric for the system under study. When the system is considered to be a
distributed system that considers the driver as an integral part, it will
enable the drivers to contribute what they are good at while the traffic sign
recognition part will present the information based on the signs detected. Furthermore, the other surrounding sensors
can also influence what is being presented.
That coordination will help to ensure that the driving takes place in
the most careful manner thereby improving the road safety.
The ability to achieve accuracy in
various environmental conditions leads to the second goal of a traffic
detection system, which is robustness.
Because it is difficult to predict the types of conditions the car will
encounter accurately, this new system should have the capability of achieving
desirable results irrespective of the environmental conditions in which the
vehicle is. That means that this algorithm should achieve integrity and
consistency under nominal condition.
The driver assistance traffic detection system should also have to be
fast. For the system to work as desired
and help to achieve the goal as to why it was designed, it should work in
real-time. Also, the detection of
traffic signs is just a small task in the whole autonomous system network. Also, in the case of autonomous vehicles,
they move at high speed and mainly require a reaction towards the traffic signs
within only a few seconds.
Lastly, the fourth evaluation metric of
a vision-based driver assistance algorithm is the cost. There are various
sensors requires including GPS, radar, Lidar, and inertial sensors, to achieve
the desired system for the car in question.
All these require a lot of spending, but the designers should make sure
that too much expense should not go to the same. The industry also should achieve a relatively
high detection rate with the use of low-cost cameras, even though the camera is
not the most expensive item in our case.
It is prudent that when the various subsystems and sensors are being
purchased the cost be put into consideration since for any investment, the goal
is to minimize the costs and maximize the benefits. Therefore, this useful
metric will be important in evaluating the success of the system under
development.
1.2.
Statement of Purpose
A vision-based driver assistant system
can detect the road signs based on their color and shape. These two areas, the
shape, and the color, are crucial in implementing a driver assistant system
that can be effective and reduce the road accidents even in an autonomous
vehicle. Many systems have implemented
the vision-based driver assistant systems to help in detecting the road signs,
and hence the purpose of this paper is to design and implement a vision-based
driver assistant system that can detect the signboards. For instance, at a
junction, it will be easy for a car to know the right direction that it will
take to ensure that it gets to the desired destination.
1.3.
Statement and Importance of Problem
In the recent years, the vision-based
detection systems for assisting drivers have been incorporated in the top-of-the-line
models by various manufacturers. However, a more effective sign detection
solution has yet to materialize. While many such systems involve the detection
of the normal road signs like the traffic lights from other vehicles, the marks
on the roads, and identifying the distance between one car and the other so
that an accident can be avoided, they have not effectively included the issue
of the signboards. A more general
vision-based sign detection systems for driver assistant ought to incorporate
every aspect of the road to make sure that the system not only helps to avoid
the road accidents, but it also helps in making the right judgment for the
direction to move particularly in the case where there are many diversions.
The computational power and the
flexibility of the vision-based driver assistant systems have recently
increased due to the development of system-on-chip technologies as well as the
advanced computing power of new multimedia devices. It is thus crucial to take advantage of the
advancement of technology to design powerful systems that can help to make life
easier than in the past. The issue of introducing the driver assistant
technologies onto the roads has been slow due to the lack of a wide coverage of
all the areas that should be covered.
This research will thus add a fundamental ingredient to the solution of
vision-based driver assistant systems. The introduction of the signboard
detection into this big picture will present a crucial success into the
vision-based driver assistant systems. When a driver approaches a junction,
where there are many directions to take, signboard detection system will be a
very important facet to determine the right direction to take.
1.4.
Research Objective
The core purpose of this dissertation is
to come up with an algorithm that detects the signboards on the road and embed
that in a vision-base driver assistant system.
It will also ensure that driver assistant vehicles will receive an alert
in real-time for any signboard detected on the road and then make a decision on
which road or direction to take to arrive at the desired destination.
1.5.
Experimental Approach
In this paper, there will be a design of
architecture for the vision-based driver assistant system that is based on an
image processing technology. A camera will be mounted on the car’s windshield
to determine the roads’ layout as well as the vehicle’s position on the lane
and then detect the signboard to determine the lane the vehicle should
take. The resulting image sequence is
then analyzed and then processed by this system that will then automatically
determine and report the possible directions for the vehicle to take.
1.6.
Significance of the Study
One of the unique abilities of this
study is that it offers an affordable solution to the vision detection driver
assistant system with the use of the available image processing technology.
This approach will especially benefit the designers of driver assistant system
and the ongoing projects of autonomous vehicles. This new system also provides
real-time information based on the signboard images processed; it is cost
effective, accurate and consistent as compared to the other expensive Matlab
algorithms. The system is a great improvement to the vision-based driver
assistance systems since brings novel ideas especially in the area of reading
the information on the signboards instead of just reading the traffic signs
like in many systems. It forms a vital foundation on which future systems can
be built to ensure that the entire road environment is covered to help the
driver with all the necessary information to make right decisions and avoid
accidents.
1.7.
Limitations
a. The
new system only concentrates on the detection of signboards and not the other
road signs like the vehicles in front and behind. It also does not include such
signs as the warning lights or even a closing pedestrian.
b. The
system can only detect the signboards that are not obscured and at a closer
distance, not the ones that are raised too high.
c. In
case the signboard is too close, the system may not help the vehicle to make a
quick decision on where to turn, especially if the vehicle was moving at high
speed.
d. The
system will need to be enhanced so that it can be used in the autonomous
vehicles.
e. The
system can also have difficulties in perceiving the signboards or other objects
during the night as well as in extreme weather conditions.
f. The
system cannot warn the driver of other dangers on the road beside the
information on the distance and the signboards that are ahead.
CHAPTER II
REVIEW OF LITERATURE
Previous studies on vision-based driver
assistance systems have attempted to identify other vehicles, traffic signs,
obstacles, and pedestrians in on-road traffic scenes by capturing image
sequences with the use of image processing as well as pattern recognition
techniques. The adoption of various concepts and definitions of the objects of
interest on the road, the application of various techniques that can capture
image sequences is useful in detecting obstacles and other signs on the road.
Many previous studies have focused on detecting objects and searching for
specific patterns on the detected images.
Many techniques have been used in detecting, classification, and
interpretation of the data collected on the images. Many different studies use
different techniques in all these stages of image processing, and these various
studies are analyzed in this chapter of the dissertation.
Vision-based driver assistant systems
have the capability of detecting road signs based on their shape and color.
That is why in literature there are majorly two methods to solve the issue of
traffic sign recognition, and these include segmentation, with the use of color
information, or the analysis of the edges of grey-scale images obtained from
the camera. Many sign recognition
systems in the past have been focused on these two systems. Those for sign detection systems have been
based on the normal traffic signs like the one for a corner, bumps, zebra
crossing, stiff cliff among other road signs besides the road. Viola and Jose
(2001) developed a vision based detector based on machine-learning that uses an
attentional cascade consisting of boosted Haar-like classifiers.
Viola and Jose’s visual sign detection
framework is capable of processing sign and objects images extremely quickly
while ensuring a high detection rate. The first key contribution in their
system is the introduction of an image representation known as “Integral Image”
that allows the computation of features utilized by their detector to be done
very quickly. The other key contribution is a learning algorithm whose basis is
AdaBoost that selects a few crucial visual features and provides extremely efficient
classifiers. The other vital contribution is a technique for combining the
classifiers in a “cascade” to allow for the image’s background regions to be
rapidly discarded so that more computation can be done on promising object-like
regions. This system provides face detection performance that can be compared
to the best vision-based sign detection driver assistant systems. However, this system usually concentrates on
detecting the signs that are in the form of objects and do not contain letters
or numbers.
Another work in the vision-based sign
detection for driver assistance is the work of Huang and his colleagues (2002)
who came up with a Gaussian filter that is a peak-finding procedure as well as
a line-segment grouping procedure that is capable of detecting lane marks. The vehicle detection is then achieved by
leveraging the feature of undersides, symmetry properties, and vertical edges.
Sun, Bebis, and Miller (2002) also proposed a car feature extraction as well as
a classification technique that uses a support vector machine in combination
with a Gabor filter. These authors used
a generic algorithm for the purpose of optimizing banks while clustering was
also used to find the filters that contain similar parameters and deletes
redundant filters. The concern with
these two systems is that the data set collection and the iterative training of
the classifiers are complex even if they are performed in advance.
The other work that uses edge-analysis
with a grey-scale image is that of Gavrila (1999) where he uses a
template-based correlation technique to detect potential traffic signs embedded
in images. The technique involves the ‘distance transforms’ method whereby you
start with an edge image and then perform a matching with the template of the
searched signs. It organizes the
templates hierarchically to reduce the number of possible operations. The problem with this method is that it
entails a high computational cost to come up with a real-time system. Another remarkable work is that of Barnes and
Zelinsky (2004) where they used the Hough Transform variation. Loy and Zelinsky
(2003) based this system on an earlier system, which is a quick method for
detecting points of interest with the use of a system comprising of radial
symmetry. The information used is of a
magnitude, and the phase of the gradient read from the grey-scale
edge-image.
Even though Loy and Zelinsky’s method
cannot detect only the circular signs, it was improved to detect even
rectangular, square, and even octagonal signs in Loy and Barnes (2004). A
self-organizing map can be leveraged to enable the extraction of contours and
recognize the shapes of different traffic signs. The histograms of oriented gradient have also
been used to filter the road signs from other signposts on the roadsides. It is
a useful system that forms that basis for this research because a system is
required that can differentiate the other objects that are not the targeted
regions of interest. The researcher will also extract the knowledge on contour
extraction from this system to help in extracting the contours of the objects
detected. The limitation of the system above is that it may not work
effectively in extreme weather conditions and also during the night.
The resulting image from the detected
sign has to be analyzed using a classifier to determine whether the detected
candidate regions are real traffic signs or not. This stage is known as classification. The most commonly used tools to achieve that
area the neural networks in their various topologies (Garvrila 1999; de la
Escalera et al., 2004; & Broggi, et al., 2007). A normalized image of the
potential traffic sign is used as an input vector. Even though the neural
networks comprise the key tool used in this stage, it is not the only one that
can be used. The template-matching methods can also be used. In the template
matching technique, a normalized cross-correlation between the possible traffic
signs and the templates stored in a database is used. Also, in García-Garrido et al. (2012), a new approach
that uses a support vector machine with a Gaussian kernel is leveraged to
accomplish the classification stage.
Buciu, Gacsadi, and Grava (2010) also
propose a system that is capable of monitoring the state of the driver. The
researcher says that drowsy driving is a considerable problem that can result
in thousands of automotive accidents each year. The researcher gives the
statistics of France where about 30 percent of car crashes occur every year,
and they are responsible for about one-third of the fatal crashes that occur on
the French highways. The driver state monitoring is something that started
about 37 years ago, and they are still very active. These driver’s state
monitoring systems should have the ability to detect drowsiness via the
vehicle’s behavior, detect the drowsiness of the driver via the driver’s
physical behavior, or via the driver’s physiological behavior. The research in
this paper can draw ideas in the future on how to incorporate machine learning
into the vision-based driver assistance system and come up with a system that
is more robust and intelligent upon this useful research. The only problem is
how to be able to detect the physiological behavior or the driver, but it is a
feasible system.
CHAPTER III
METHODOLOGY
3.1
Introduction
The development of a vision-based driver
assistance system is imperative in the context of the road conditions. This study aims at developing a system that
can detect signboards on the road and alert the driver promptly so that he/she
can make the correct choice on the direction he/she should take to arrive at
where he/she wants to go. Such a system aids the driver in making the correct
turns based on the destination that the driver wants to take. That means that
the system should be accompanied with a pre-determined data on the route and
destination of the vehicle. This project
will follow some steps that are aimed at ensuring that a complete system is
designed as anticipated. The road
signposts are usually confusing, and the new road users find it problematic to
make the correct turns especially where there are more than two possible routes
to take from a given point.
The proposed vision-based system
integrates an effective vision-based as well as processing modules such as
signboard detection, event recording functionalities, and collision warning
features because to avoid colliding with another vehicle when negotiating a
corner or when changing the lane to enter the required one. These functioned will be implemented to identify
the target signboards, determine the vehicles in front of the car, estimate
their distances, and the record the traffic event videos. In the following subsections, these
processing modules are described in detail.
3.1.1 Segmentation.
The essence of this segmentation step is
to have a rough idea concerning the signs that might be on the signboards
thereby narrowing the search space for the subsequent steps. A unique approach has been proposed: leveraging
a biologically based attention system, it gives a heat map that shows the areas
on a signboard or signpost where the sign is likely to be found. These input images, in this case, have to be
capture from the camera, also referred to as the vision system. Those sensed frames are the road environments
that appear on the camera that attached in front of the host car. The task of this object segmentation section
is to extract an object from the road environment to facilitate the rule-based
analysis of the object. To reduce the
costs of computation of extracting the objects, this module first has to
extract a grayscale image to determine if it is just an object or if it is
signboard.
To extract those images on the road from
a given gray intensity image, pixels of objects have to be separated from the
other object pixels that have different illuminations. Since most of the
signboards that are above the road are green, it will be easy first to separate
the images on the vision system based on color. Therefore, it is imperative for
an effective thresholding technique that can automatically determine the
suitable number of thresholds for segmenting the object revisions from the
detected image. An effective multilevel
thresholding technique has been proposed in this paper to ensure fast region segmentation. This technique automatically decomposes
captured road scene images to produce homogenous threshold images using a
discriminate analysis concept.
3.1.2 Spatial analysis and clustering.
To identify the potential
signboard-light components after completing the object segmentation, a
component extraction process is performed on the object plane to locate the
linked components of the objects. The process tries to identify the rectangular
shaped green object above the virtual horizon of the y-axis. This virtual horizon is at a distance of
above 150 centimeters above the road surface since many signboards are placed
at this distance above the road surface.
Thus, a clustering process has to be used to the components to cluster
them into many meaningful groups. This
group may consist of traffic lights, road signs, as well as other illuminate
objects that have been raised above the ground.
The image identification process then processes those groups to identify
the actual signboards.
To preliminarily screen out the
non-signboard objects like the street lamps or traffic lights, the objects that
appear below one-third of the virtual Y-axis horizon are filtered out (that is,
only the components located above the constraint line). That is because this study assumes that the
signboards are placed above the road surface.
Also, to determine the direction that
the sign on the signboard is pointing to, it is vital to identify the
components at the head side and the tail site before carrying out the response
analyses. The usefully distinguishable feature of the tail and head sides of a
road sign is the arrow that automatically represents the forward direction.
When there is an object that is close to the camera-assisted car, there are
blooming effects in the CCD cameras that may hinder the camera from focusing on
the signboards above. Thus, the vehicle
should be at the required distance from any obstruction, especially a tall
vehicle containing some goods that are likely to obscure the camera. After the
identification, the objects are then merged and clustered into rectangular
green component groups if they have signs denoting an arrow on them, are close
to one another, and are aligned.
3.1.3 Sign tracking and identification phase.
These techniques obtain the image groups
of the potential signboards in each captured frame. However, because adequate features of some
potential signboards may not be immediately captured from the single image
frames, there is a need for a tracking procedure that can analyze the information
of the possible signboards from successive image frames. The tracking
information is then used to refine the detection results and suppressing the
errors that may have been caused during the object segmentation process as well
as the spatial clustering process. The tracking information can also be useful
in determining the useful direction of the signpost, the lane to take, and
other useful information as some lanes.
To distinguish the real signboards in
each frame, the proposed system applies a rule-based process to each potential
tracked signboard to determine whether it includes actual signs on it or other
illuminated objects. Of course, the signs should also be sidelined with the
potential information to show the destination points of each.
3.2.
Signboard Distance Estimation Module
For estimating the distance between the
host car’s camera and the detected signboard, the proposed systems applies the
perspective of range estimation of the CCD camera that was introduced by Stein,
Mano, Shashua (2003). The origin of virtual signboard coordinate appears at the
center of the camera lens. In the same manner, the X and the Y coordinate axes
of this signboard are parallel to the same coordinates of the capture images,
and the Z-axis appears along the optical axis and is perpendicular to the plane
that is formed by the horizontal and vertical lines (X and Y coordinates). A target signboard on the road that appears
at a distance Z in front of the car projects to the camera’s image at the
y-coordinate. Therefore, the
single-camera range estimation perspective can be useful in estimating the
Z-distance between the camera-assisted car and the signboard using the below
equation.
Z = k. ((f.H)/y)
With k being the factor for converting
the distance from pixels to millimeters for the mounted camera at the height H,
and f being the focal length measured in meters.
3.3.
Research Design
The design that is used in this research
is the experimental design. A CCD camera is selected, and a place for mounting the
camera is determined. This point of attachment of the CCCD camera on the host
car should be in such a way that it can view the signboards that are usually
raised to some distance above the road surface. Being able to focus on that
point will help only to view these signboards that are also usually of a
specific color depending on the country. Different countries have different
colors for the road signboards. The most commonly used color for the same is
green. Also, the computer system is used to distinguish between the actual
signboards and other boards or objects that may be placed on the roads such as
advertisements and caution messages.
After the camera detects the object
above the road surface, the object must be filtered using a rule-based approach
to determine if it an actual signboard of it is another object. An arrow is
then searched on the object because a signboard has to contain some arrow
showing which lane leads to where so that the car can turn to the desired
direction. The information is given in
real-time allow for making the correct decisions to be made at the right time.
Thus, the distance of the car should also be calculated how it is from the
signboard and at what point should it change lanes if necessary.
3.4.
Data Collection, Subject Selection, and Description
The experimental car is given by NY
Toyota car dealers while the CCD cameras are acquired from the Princeton
Instruments. The JPEG files also used to
display the images in 2-dimensional view while the STL files present the images
in 3-dimensional. The features of the
signboard are extracted and then recognized by the image processing system
after which the tracking process is done based on the optical flow to reduce
the complexity of computing.
3.5.
Procedure
i.
First of all the frame is captured by
the camera after that focuses on any object that is raised above the road
surface.
ii.
The region of interest is then extracted
from the captured images. This region of interest is the arrow that is usually
found on the signboard showing the direction to where different lanes lead.
iii.
The computing system then reads the
identified point at the center of the image which has been determined to be a
real signboard. The head and the tail
parts of the arrow have to be examined and interpreted to determine the
direction of each lane.
iv.
The distance of the signboard from the
car is then determined to find out at what point the car should turn or change
the lane to get to the targeted destination.
v.
A tracking process then follows to
determine if the capturing and processing have been done correctly, after which
the computer system helps the driver to make the final decision.
To effectively assess the obtained
results, a statistical hypothesis was developed to find out if the vision-based
driver assistant system is more effective as compared to the manual means of
watching for the signboards and then making the correct decision. As part of this analysis, below are the
hypotheses that were developed.
H0: The vision-based driver assistant
system of signpost recognition is more effective in detecting the signboards on
the road and determining the lane to take in a junction with multiple lanes.
H1: The use of the vision-base driver
assistant system for detecting the signboards on the road is not as effective
as the manual method of detecting the signboards and making the correct
decisions.
3.6.
Limitations
The one main limitation of this study is
that the cameras may not work well during the night in case there is no
lighting on the signboards. Also, it may be hard for the camera to perceive the
signboards if the camera is mounted on a small car and the car is too close to
a tall vehicle that is also carrying some goods that may obscure the
camera. The other limitation of the
study is that the camera may not work well when it is raining heavily. When the
car is also moving at a fast pace, and the driver does not know the road very
well, he/she may make a wrong turn off a change of lane especially if they read
the information from the system very late or the results are not displayed in
real-time.
CHAPTER IV
RESULTS AND DISCUSSION
4.1.
Results
The experiments were performed on the
highway during daytime with sufficient light. The results presented here were
obtained from the image sequences captured by the camera at different points on
an 80-kilometer road. The performance of
the entire system was computed over a test set of 60 stereo pair of images that
had a resolution of 320 X 240 that corresponds to the 80-kilometer road. The results were obtained by taking into
account that the signboards detected were positive samples denoted by P whereas
the negative samples represented by letter N were the noisy objects detected
and thus no signboards. Therefore, every signboard can be detected of
classified as being a true positive TP or a negative, positive FN. A TP occurs when the prediction and the real
values are positive while the TN occurs when the prediction outcome and the
real values are negative. When the
prediction result is a negative while, and the actual values are positive, the
outcome is a false negative.
To make the proposed system operate as
anticipated, the CCD cameras receiving input sequences of the objects was
mounted onto the windscreen immediately behind the windshield inside the car
with adequate height to allow the capturing of the appropriate regions for
encompassing the interesting signboards to be monitored as shown in figure 1
below. The view angle of the camera was then calibrated to make it parallel to
the road surface with an elevation angle of 60 degrees, whereas the focal
length of the camera was set as 20mm. The peripheral devices such as the image
grabbing devices, in-car control devices, and the mobile communication systems
were also included in the embedded platform to accomplish the internal
vision-base driver and surveillance system.
The system was then tested on many videos of real road scenes under
different conditions.
system for detecting signboards on the road. As the figure shows, the system
consists of three buttons for system configuration, starting, and
stopping. The function of the system
configuration button is to set the system values like the voice volume, traffic
scene video recording, and control signaling, while the system starting and
stopping buttons are used to start and stop the system respectively.
A tracked
signboard was considered classified when the output from the classifier is
greater than a given threshold for over five times. It is difficult to classify
the signboards at night particularly the circular ones that have less
variability. That was also an advantage because the circular objects had to be
eliminated, as they do not represent the shape of a signboard, which is always
rectangular. In that sense, the sensitivity of the circular objects at night
reached not more than 80 percent.
However, the sensitivity, as well as the precision of other objects,
exceeded 95 percent any lighting conditions.
One objective of
this experiment was to get a system that can work real-time. In this sense, an
average run-time was measured, and this is as shown in the following diagram,
figure 1. That consisted of the three
processes: the process of extraction, encoding, and selection contours, the
process of transformation implementation of the Hough (HT), and the tracking
and classification process denoted by SVM.
Even though the results were obtained
using an offline process, the obtained runtime of 35ms with a deviation of 19
ms can allow for real-time performance of the system. In future, a real-time
implementation is considered to confirm that this experiment can effectively
work in real-time.
For the computational time issues,
the required time for computing one input frame is determined by the complexity
of the signboard objects being captured.
Most of the time spent on computing took place during the clustering
process of the potential signboard objects.
Based on the system, the vision computing phases of the proposed driver
assistant system require 30ms averagely to process a frame of 320 X 240 pixels
while the traffic scene video recording took approximately 5ms per frame when
there is hardware acceleration. The computational cost presents help to ensure
that the proposed driver assistant system effectively satisfies the real-time
processing demand that in this case is set at ten frames per second. Therefore, this proposed system offers a
timely assistance and warning for drivers so that they can make the correct
decisions at a point where there are many possible lines to take.
4.2. Discussion
Currently, many vision-based driver
assistant systems that have been proposed or developed only involve the
detection of lanes, signs, or other objects on the road as the cars. Also, such
systems are only aimed at reducing traffic accidents. There is not a system
that has been developed to assist the drivers in identifying the correct lane
to shift in a junction where there is more than one possibility, and there are
signboards above the highway. In such a case, it is usually problematic for the
new drivers and other drivers who may be tired or those who may be driving at
very fast speeds because they may not make the correct decisions thereby ending
in the wrong lanes. That is what this system addresses to bridge the gap that
exists.
From the results, we can effectively
determine which hypothesis we should take, whether it is H0 or H1. The
effectiveness of the system thus makes it obvious to take the null hypothesis.
The system that has been developed in this paper shows that the vision-based
driver assistant system produces timely results that can help the types of
drivers identified above to make the correct decision when on a highway and
there are more than one exit lanes from the main highway. It was observed that
the CCD camera is very effective when used in the proposed system because it
produces high-quality images that make it easy for the computing system to
determine if the object is a real signboard or it is a different object. These high-resolution images captured by the
CCD camera also make it easy to identify the signs represented by the arrows on
the captured objects to determine if they are actual signboards or not.
The type of images that were captured
by the CCD camera was the JPEG images and the frames measured 320 by 240. For a
small screen like the one that was used by the system, this can be a good
resolution. Other researchers have preferred to capture images and present them
as VGA images, but all depends on the display system that is being used to
display the images. It is also worthy to note that the system and even the
images require portability. The JPEG images captured in this experiment can be
portable because almost any type of screen including the LED and the LCD screens
can display these images. The JPEG images are also not computationally
intensive; hence, this system is cost effective in that area. That is why these
images were processed very first by the system.
The camera was able to capture the
objects that were raised to some distance above the road surface. Many times the camera was able to perceive
the signboards because there were no obstructions such as big vehicles carrying
tall heaps of luggage. The car had to reach to a certain point where the angle
would be 60 percent to the signboard and then capture the image. Sometimes when
negotiating a corner, there were some images being captured at the same angle
the system would only report those images after successfully processing them
and determining if they are actual signboards. With this kind of intelligence,
the system can help distracting the driver when the camera captures the other
images that are not the actual signboards.
The user interface of the system also
helped largely to manipulate it and use only when there was a need to do so.
The system’s user interface consisted of the system starting, system stopping,
and the system configuration buttons. The system’s start button helps the
driver start using this intelligent system while the stopping button helps the
driver to turn the system off when he/she does not want to use it. However,
since the environment for testing the system was on a highway, the system was
not turned off throughout since it was required to be used. Also, the driving
speed of 60km/h helped the system to capture and process the images as desired
effectively. At a very high speed exceeding 100km/h, the system may not be
effective although this was not tested by this experiment.
From the experiment, it is also
observed that the camera can capture up to 12 images per second and this is one
of the strengths of this system. Many systems that have been developed before
usually capture up to seven images per second and so this is a great
improvement to the vision-based driver assistant system. In many highways, it is usually hard to find
more than five signboards on the same point of the highway, but there are also
times when you may find some advertisements or road caution boards also put
alongside the signboards, thus it is vital for the system to capture as many
images as possible per second. That will also help the system to apply to any
highway, even the one that may be having several objects together with the
signboards at the same point on the highway.
A set of different objects has been
classified in the system as circular, triangular, or rectangular. This large
set of the shapes is a great improvement concerning the other works. That means
that the system can be manipulated to be used to detect other road signs that
are in various shapes including among the ones mentioned above, the diamond
shapes. Also, the testing of the experiment in the real world has helped to
ensure a high recognition rate of the signboards. It makes the system to be
reliable because it can be used the way it is without any modifications. That
also makes it different from many similar systems that are merely simulations
and not real experiments. As seen from the results, the precision and
sensitivity are above 95 percent, and it reduces only during the extreme weather
conditions as well as during a night where the signboard is not adequately
illuminated.
Additionally, using the 12V
communication system presents a novel solution to the issue of discarding those
detected objects that do not pertain to actual signboards instead of utilizing
geometrical constraint. In any research, coming up with novel ideas on how to
solve the impending problem is the main objective, and in this research, that
has already been achieved. The system can also be programmed to make it
multipurpose by introducing other functionalities like identifying the road
signs, identifying pedestrians on the road, identifying other vehicles in front
and even behind, and doing accurate readings even in extreme weather
conditions.
The vision-based technologies were
integrated and implemented using an ARM-DSP multi-core platform that includes
peripheral devices such as mobile communication and image grabbing devices.
These modules and other in-vehicle control systems were integrated so that an
in-car embedded vision-based driver assistance, as well as surveillance system,
can be accomplished. The experimental results show that this proposed system
can be effective and offer benefits for integrated signboard detection and
traffic event recording. All these factors help the driver to make desired
surveillance in different road environments as well as traffic conditions. That makes it possible also to make minimal
configurations on the system and then apply it in other instances on the
traffic detection.
CHAPTER
V
CONCLUSIONS AND RECOMMENDATIONS
5.1. Conclusions
This paper has presented a
vision-based signboard recognition system that can assist the drivers driving
on a highway to make prudent decisions. Many highways usually have several
lanes you may be required to keep on shifting from one lane to another depending
on your destination. With this system, the driver should not have to worry
since it does for him/her almost everything they need when especially they are
confronted with a dilemma on which lane to shift to, and there are signboards
above the highway. The system developed in this case can make very quick
calculations and processing of the images received from the camera that is
mounted below the windscreen just behind the windshield. One of the key strengths of the system is
that it allows the detection and processing of the images in real-time. To
receive results in a second means that you are receiving the results in
real-time and thus there is no distraction.
Also, the experiment showed that the
encoding of contours helps to solve, in many instances the issues of
bifurcations, discontinuities, and the change of direction. Thus, makes this
proposed vision-base driver assistant system for detecting signboards to be
reliable and accurate because it obtains an average detection rate of 95
percent in all the lighting conditions. It is also applicable to the
triangular, rectangular, and arrow signs and objects. The fact that this system can also filter out
other shapes that do not constitute a signboard makes it reliable. Furthermore, the detection of the signboard
in this system is adaptive because of the use of adaptive thresholds and the
application of the Hough transform based on the information the system receives
from every contour.
5.2. Recommendations
The system developed in this case
should be compared with other existing approaches for clearly offering the
baseline of improving the results obtained in the system. However, it is
impossible to make such comparison because there are no common criteria or
frameworks for evaluating the traffic sign detection systems. I recommend that
there be frameworks that can form the basis for evaluating the traffic sign and
signboard detection systems. If such frameworks are available in the future,
this research will also make sure that the comparison to find the areas of
improvement. Another area that also presents an opportunity for improvement is
the area of sensor installation. Sensors can be installed on all the signboards
to help the vision-based systems to detect them easily. That will also help to
reduce the need to install expensive systems in the vehicles and will only
require these vehicles to have less expensive sensors and transmitters. In the
future, that is what this research will focus on.
Additionally, in future, the research
will focus on using the vehicle dynamics, such as the vehicle trajectory, yaw
rate, speed changes, vehicle direction, and steering wheel position among
others. The essence of using these vehicle dynamics is to improve the
robustness of the process of discarding the unwanted traffic signs or objects.
Automatic traffic sign recognition for a driver assistant system can be very
important, although it also embodies other possible applications like the
tracking of the inventory system of the traffic signs and automatic inspection
of these signs to provide a safer response as well as a better maintenance
signposting. Such an automatic system can also help in building as well as
maintaining of the maps of road signboards and other traffic signs. All these
applications will entail a challenging research work for the future.
In the future, the other improvement
and extension that can be made on this vision-based driver assistance system
are the integration of some complex machine learning techniques like Support
Vector Machine classifiers on many cues such as on the car lights and bodies.
That will help to further enhance the signboard and other traffic signs’
detection feasibility under night and difficult weather conditions. Also, it
will also improve the classification capability on those vehicle types that are
more comprehensive like the sedans, Lorries, buses, trucks, and motorbikes. It
is also likely that the research in other areas like lane detection can be of
great benefit here. Additionally, the idea is concerning the surrounding; the
connection between the knowledge of the weather and the lighting conditions at
a given time can enhance the robustness of the system. Otherwise, such a system can be more useful
during the night compared to during the day.
Barnes, N. & Zelinsky, A.
(2004). Real-time radial symmetry for speed sign detection. Proceedings of the
Intelligent Vehicles Symposium; Parma, Italy. 14–17 June 2004; pp. 566–571.
Broggi, A., Cerri, P., Medici, P.,
Porta, P. P., & Ghisio, G. (2007, June). Real time road signs recognition.
In Intelligent Vehicles
Symposium, 2007 IEEE (pp.
981-986). IEEE.
Buciu, I., Gacsádi, A., & Grava, C. (2010). Vision-based
approaches for driver assistance systems. Proc. ICAI'10, 92-97.
De La Escalera, A., Armingol, J.
M., Pastor, J. M., & Rodríguez, F. J. (2004). Visual sign information
extraction and identification by deformable models for intelligent vehicles. IEEE transactions on intelligent
transportation systems, 5(2),
57-68.
García-Garrido, M. A., Ocana, M.,
Llorca, D. F., Arroyo, E., Pozuelo, J., & Gavilán, M. (2012). Complete
vision-based traffic sign recognition supported by an I2V communication system. Sensors, 12(2), 1148-1169.
Gavrila,
D. (1999). Traffic sign recognition revisited. Proceedings of the DAGM-Symposium; Bonn, Germany. 15–17 September 1999; pp. 86–93.
Huang, S. S., Chen, C. J., Hsiao,
P. Y., & Fu, L. C. (2004, April). On-board vision system for lane
recognition and front-vehicle detection to enhance driver's awareness. In Robotics and Automation, 2004.
Proceedings. ICRA'04. 2004 IEEE International Conference on (Vol. 3, pp. 2456-2461). IEEE.
Loy,
G. & Zelinsky, A. (2003). Fast radial symmetry for detecting points of
interest. IEEE Trans. Pattern
Analysis for Machine Intelligence, 25, 959–973.
Loy, G., & Barnes, N. (2004,
September). Fast shape-based road sign detection for a driver assistance
system. In Intelligent Robots
and Systems, 2004.(IROS 2004). Proceedings. 2004 IEEE/RSJ International
Conference on (Vol. 1, pp.
70-75). IEEE.
Mogelmose, A., Trivedi, M. M.,
& Moeslund, T. B. (2012). Vision-based traffic sign detection and analysis
for intelligent driver assistance systems: Perspectives and survey. IEEE Transactions on Intelligent
Transportation Systems, 13(4),
1484-1497.
Morris, B., & Trivedi, M.
(2010, June). Vehicle iconic surround observer: Visualization platform for
intelligent driver support applications. In Intelligent
Vehicles Symposium (IV), 2010 IEEE (pp.
168-173). IEEE.
Stein, G.P., Mano, O., &
Shashua, A. (2003). Vision-based ACC with a single camera: Bounds on range and
range rate accuracy. Proceedings of IEEE
Intelligence Vehicle Symposium, 2003:120–125.
Sun, Z., Bebis, G., & Miller,
R. (2002). On-road vehicle detection using Gabor filters and support vector
machines. In Digital Signal
Processing, 2002. DSP 2002. 2002 14th International Conference on (Vol. 2, pp. 1019-1022). IEEE.
Trivedi, M. M., Gandhi, T., &
McCall, J. (2007). Looking-in and looking-out of a vehicle:
Computer-vision-based enhanced vehicle safety. IEEE Transactions on Intelligent
Transportation Systems, 8(1),
108-120.
Viola, P. & Jones, M. (2001). Robust real-time
object detection. International Journal of Computer Visual, 57(2), 137–154.
Sherry Roberts is the author of this paper. A senior editor at MeldaResearch.Com in Write My Essay Today services. If you need a similar paper you can place your order from pay for research paper services.
No comments:
Post a Comment