Applications of Image Processing and Real-Time embedded Systems in Autonomous Cars: A Short Review

Vidya Viswanathan & Rania Hussein
International Journal of Image Processing (IJIP), Volume (11) : Issue (2) : 2017 35
Applications of Image Processing and Real-Time embedded
Systems in Autonomous Cars: A Short Review
Vidya Viswanathan vidyakv@uw.edu
Department of Electrical Engineering
University of Washington
Bothell, 98011, USA
Rania Hussein rhussein@uw.edu
Department of Electrical Engineering
University of Washington
Bothell, 98011, USA
Abstract
As many of the latest technologists have predicted, Self-driving autonomous cars are going to be
the future in the transportation sector. Many of the billion dollar companies including Google,
Uber, Apple, NVIDIA, and Tesla are pioneering in this field to invent fully autonomous vehicles.
This paper presents a literature review on some of the important segments in an autonomous
vehicle development arena which touches real time embedded systems applications. This paper
surveyed research papers on the technologies used in autonomous vehicles which includes lane
detection, traffic signal identification, and speed bump detection. The paper focuses on the
significance of image processing and real time embedded systems in driving the automotive
industry towards autonomy and high security pathways.
Keywords: Image Processing, Hough Transform, Canny Edge, Open CV, Polygonal
Approximation, Computer Vision.
1. INTRODUCTION
Image processing is one of the main drivers of automation, security and safety related application
of the electronic industry. Most image-processing technologies involve several steps like treat the
image as a two dimensional signal and apply standard signal processing techniques to it. Images
are also handled as 3D signals where the third dimension is the time or the z-axis. Highly efficient,
low memory and reliable solutions can be achieved by utilizing Embedded Systems and Image
processing to bring out the benefits of both for applications.
Google is one of the billion dollar companies who has demonstrated its own driverless car, a
design that does away with all conventional controls including the steering wheel, and other
astonishing technologies. In their driverless car, Google has not only included Image Processing,
but also many other amazing technologies and one of the most important among them is Lidar,
which stands for “Light Detection and Ranging”. It consists of a cone or puck-shaped device that
projects lasers which bounce off objects to create a high-resolution map of the environment in
real time. In addition to helping driverless cars to “see”, Lidar is used to create fast, accurate 3D
scans of landscapes, buildings, cultural heritage sites and foliage. Some of the other technologies
include Bumper Mounted Radar for collision avoidance, Aerial that reads precise geo-location,
Ultrasonic sensors on rear wheels which detects and avoids obstacles, software which is
programmed to interpret common road signs etc. Apart from these, there are altimeters,
gyroscopes, and tachymeters that determine the very precise position of the car and offers highly
accurate data for the car to operate safely. The synergistic combining of sensors is one of the
most important factors in this autonomous car which includes the data gathered altogether by

these sensors are collated and interpreted by the car’s CPU or in built software system to create
a safe driving experience.
Apart from Google, many other companies like Tesla, Audi, Uber have also developed their own
driverless cars and have tested potentially.
This paper concentrates on how Image processing can be used in vehicles to drive the automotive
industry to completely autonomous and high security pathways. A real time embedded system
environment is inevitable in an automotive application. Also, the scale of the industry is very high
so the solutions should be cost efficient, fast and reliable. This paper intends to highlight these key
points.
2. AUTONOMOUS VEHICLE CONTROL USING IMAGE PROCESSING
In autonomous vehicles, the driving commands from a human driver are replaced by a controller
or a microcomputer system that generates these commands from the information it gets as it’s
input. Since this paper deals with the applications of image processing in autonomous control of a
vehicle, the input given to the microcomputer system is the visual information obtained from a
camera mounted on the vehicle. This section explains in detail, some of the important factors in
Autonomous vehicles such as Lane Detection, Traffic Sign Detection, Speed Bump Detection,
Steer By Wire System etc., which uses the processing of received image inputs and the
algorithms used in them. Lane Detection represents a robust and real time detection of road lane
markers using the concept of Hough transform in which the edge detection is implemented using
the Canny edge detection technique and Spiking neural network technique. Traffic Sign Detection
includes the recognition of road traffic signs in which, the concept of Polynomial approximation of
digital curves is used in the detection module. Speed Bump Detection represents the detection
and recognition of speed bumps present in the road which alerts the vehicle to control the speed
automatically and Steer-By-Wire system represents an autonomous steering system using an
Electronic Power Steering (EPS) module.
2.1 Lane Detection
Lane detection is one of the main part in the self-driving car algorithm development. On board
cameras are kept in and around the cars to capture images of road and surrounding of the car in
real time [1]. When the vehicle appears to deviate from the lane or vehicle safety distance is too
small, it can timely alert the driver to avoid dangerous situations.
The basic concept of lane detection is that, from the image of the road, the on-board controller
should understand the limits of the lane and should warn the driver when the vehicle is moving
closer to the lanes. In an autonomous car, lane detection is important to keep the vehicle in the
middle of the lane, at all-time, other than while changing lanes. Lane departure warning systems
have already crawled into most of the high-end passenger cars currently in market.
A typical lane detection algorithm can be split into simple steps:
1. Select the ROI (Region of Interest)
2. Image Preprocessing (gray range/image noise subtraction)
3. Get the edge information from the image (edge detection)
4. Hough Transform (or other algorithms) to decide the lane markings
Step 1: Select the ROI
The images collected by the on-board camera are color images. Each pixel in the image is made
up of R, G, and B three color components, which contains large amount of information.
Processing these images directly makes the algorithm consume a lot of time. A better idea for this
problem is to select the region of interest (ROI) from the original image by concentrating on just
the region of the image that interests us namely the region where the lane lines are generally

present. Processing on only the ROI can greatly reduce the time of algorithm and improve the
running speed. The original image is shown in Figure 1 the ROI region is shown in Figure 2.
FIGURE 1: Original Image [1].
FIGURE 2: ROI area [1].
Step 2: Image Preprocessing
Most of the road images have a lot of noise associated. So, before we do any image processing
steps we need to remove those noises. This is typically done through Image Preprocessing. Image
preprocessing includes grayscale conversion of color image, gray stretch, median filter to
eliminate the image noise and other interference information. Gray stretch can increase the
contrast between the lane and the road, which makes the lane lines more prominent. Equation (1)
represents the function which is to be applied to an RGB image to convert it to Gray Scale.
L(x,y) = 0.21 R(x,y) + 0.72G(x,y) + 0.07 B(x,y) (1)
Where
R - Red component of the image
G - Green component of the image
B - Blue component of the image
X,y - position of a pixel
Figure 3 shows the gray scaled conversion of the ROI region and Figure 4 shows the gray stretch
image.
FIGURE 3: RGB to Gray conversion of Original Image [1].

FIGURE 4: Gray Stretch Image [1].
The method of image filtering includes the frequency domain filtering and the spatial domain
filtering [2]. Spatial domain filtering is simpler and faster than the frequency domain filtering.
Spatial domain filtering can remove the salt and pepper noise from the original image and
preserve the edge details of the image. Its main principle is to use the middle value of every pixel
in the neighborhood of one pixel instead of current pixel value. The image after median filtering is
shown in Figure 5.
FIGURE 5: Median filtered image [1].
Step 3: Edge Detection
The next step is to perform edge detection on the output of the preprocessing. It is basically to
detect the lines around the objects in the images. One of the common methods of edge detection
is called Canny Edge Detection introduced by John F Canny, University of California, Berkeley in
1986. It basically uses multiple steps including Gaussian filters, intensity gradient changes to
determine edges [3]. Figure 6 shows the Canny edge detected output of the grayscale converted
image.
FIGURE 6: Canny Edge Detection Output image [1].
In recent researches, one of the main goals are to develop higher efficient edge detection
algorithm for better detecting edges from varying image quality. For that purpose, an alternative
approach to edge detection called Spiking Neural Network [1] is used, which is claimed to be a
much efficient method than canny edge detector. Figure 7 shows the sparking neural network
output of the gray scale converted image.

FIGURE 7: Spiking Neural Network Output Image [1].
Step 4: Hough Transform
Hough transform uses simple mathematical model of a straight line in a two-coordinate system, to
predict straight lines or other simple shapes from an image. Equation (2) represents the basis for
performing Hough transform.
r = x cos θ + y sin θ (2)
From the edge detection algorithm, we get a set of points which are considered part of the lane
markings, Hough transform is used to generate a line through these points from the image. It
connects points in the 2 dimensional polar coordinates of r and θ (theta), and generates multiple
lines from each point. After generating all the different combination of lines of r and θ (theta) from
each point, there will be one specific r and θ (theta) which intersects which is common to all the
lines generated. This line will pass through all the points of interest. Thus, a line is generated.
Figure 8 shows the lane detection output after performing Hough transform.
FIGURE 8: Lane detection output after Hough transform [1].
Hough transform is a widely-used method in the lane detection. It has very good suppression of
noise performance and is not sensitive to the target which is partially occluded and covered in the
image. But, because of the complexity of Hough transform, the computation speed is very slow,
and the false detection rate is large. It cannot meet the real-time requirements accurately. The
detection speed is improved by limiting the polar angle and the polar diameter in Hough transform
space as described in Q. Wang [4]. Z.L.Jin [5] detected the lane lines by finding the best
threshold of the image. Random sample consensus (RANSAC) algorithm is used to detect the
lane lines in J.Guo [6], but it has high wrong rate. J.R.Cai [7] used the peak point inverse
transform to extract line segments in Hough transforms space. All the methods mentioned above
are modified in the process of Hough transform, but do not consider the appropriate processing
methods in the image before applying Hough transform, to improve the detection accuracy and
shorten the calculation time. X.Li [1] considered all these discrepancies and integrates a set of
image processing techniques and spiking neural network to improve accuracy of the lane
detection. Initially, the ROI is set on the image captured from the on-board camera. This can
greatly reduce the running time of the algorithm and improving the detection speed; Secondly, a
set of image preprocessing techniques is applied to the ROI region, such as transform from RGB
image to grayscale, gray stretch and median filtering. Edge detection based on the spiking neural
network with biological mechanism is used to extract efficient edge lines. At last, the Hough
transform is used to detect the lane. The new ideas are that the ROI preprocessing is used to
speed up the detection time and spiking neural network is used to obtain better edge lines to
improve the detection accuracy.

2.2 Traffic Sign Detection
Traffic sign detection is key to semi and fully autonomous vehicles. The purpose of the technique
is to identify the sign boards on the side of the road and identify what they imply and let the car
computer know. Figure 9 shows the block diagram for the implementation criteria of the system.
The system has two components, one is the preprocessing module, which is based on image
processing techniques such as color segmentation, threshold technique, Gaussian filter, canny
edge detection and contours detection [8]. The second is the Detection module based on
Polynomial Approximation of digital curves technique applied on contours to correctly identify the
signboard. It uses a similar technique of converting Image from RGB to Gray to Binary and then
finding contours. Based on the shape of the sign board, the algorithm identifies what Traffic Signal
it is.
FIGURE 9: Traffic Sign Detection System [8].
The preprocessing module takes the RGB image and converts to HSV (Hue, saturation and Value)
Color space. This can then be used for color segmentation. The images are checked for circular
and triangular shapes of Blue and Red. The detection module then identifies the shape of sign
board using the technique of Polygonal Approximation of digital curves, applied on contours,
which is based on the algorithm of Ramer-Douglas-Peucker (RDP). RDP is an algorithm for
reducing the number of points in a curve that is approximated by a series of points [9]. The
purpose of the algorithm is, given a curve composed of line segments, to find a similar curve with
fewer points. The algorithm defines 'dissimilar' based on the maximum distance between the
original curve and the simplified curve. The simplified curve consists of a subset of the points that
defined the original curve. An algorithm called Freeman Chain code, which is also behind Optical
Character Recognition (OCR), is used to identify the letters, numbers and signs inside the sign
board. Figure 10 shows the simplification of a piece-wise linear curve using RDP algorithm.
FIGURE 10: Simplifying a piece-wise linear curve using RDP algorithm [9].
The system was tested on a database containing more than 900 real images [10], with complex
natural background, to test de validity of the system in different lighting condition (day, night, fog,
rain, etc.) and in different location (city, town, nature, etc.). The results show the speed,
robustness, and reliability of the system in these different situations. Tests show that the system
implemented is very fast and good enough (detection rate 99.99%), even if it is dependent on
size of the images [8]. The program may also detect some blobs which are not real traffic signs,
but 99.98% of true traffic signs are detected. A Recognition System can be used later to
distinguish these blobs separately. To more improve the system, the techniques of multithreading

can be used, which allows parallel execution of several programs at the same time. This will
significantly improve the execution time which eventually leads to a more robust and efficient
system for real-time implementation.
The new ideas proposed in this system are that the preprocessing module has processed the
input images using image processing techniques, such as, threshold technique, gaussian filter,
canny edge detection and contours, which leads to a more accurate and reliable implementation.
The detection module to detect the traffic sign patterns, is based on Polygonal Approximation of
digital curves, implemented by the algorithm of Ramer-Douglas-Peucker and the main reason to
select this method is to reduce the computational cost in order to facilitate the real-time
implementation.
OpenCV is an open source computer vision algorithm widely used in image processing
applications [3]. Some of the widely used OpenCV function calls are defined below, which can be
related to the applications described in this survey paper.
//Declarations
//Mat is 8 bit data type of image file in OpenCV can be jpg or //png
Mat input_image, output_image
Int intGradthesh_min, intGradthresh_max, aperture_size
// convert RGB image to gray
cvtColor(input_image, output_image, CV_BGR2GRAY);
//Canny Edge conversion - it needs a min and max threshold of
//intensity gradient to be provided to identify right edges and //discard wrong edges; aperture_size
default is 3.
Canny(input_image, output_image, intGradthesh_min, intGradthresh_max, aperture_size)
2.3 Speed Bump Detection
Speed bump detection is one of the key action which the vehicle need to understand in-order to
improve user comfort, as part of the Advanced driver assistance system (ADAS) in the vehicle
[11].
FIGURE 11: Flow Chart of Speed Bump Detection [11].
It is also a safety feature for the vehicle and the driver. A similar approach as the previous system
is taken here as well. Figure 11 shows the flow chart used for the implementation criteria in this
system. First the image is converted from RGB to gray using the Gray scale function. Then the
Grayscale is converted to binary image. A morphological image processing is performed on this
binary image [11]. A morphological processing is a set of operations that transform images per the
size, shape, connectivity using the concept of set theory and other function. The process includes
three steps, Opening, Area opening and Filling. After which a spread over Horizontal projection is

performed. Projection is a one-dimensional representation of an image computed along horizontal
and vertical axis. It's the summing up of pixel value in rows and columns. The results show that
this type of speed bump detection is ideal for all the bumps with zebra crossing pattern. Figure 12
shows the analysis and results of the implementation used in this system.
FIGURE 12: Horizontal Projection width increases as the vehicle approaches a speed bump [11].
M.Jain [12] and B.R [13] gives an overview for detection of speed bumps using accelerometer.
M.P[14] focus on the sensing component, accelerometer, microphone, GSM radio, and/or GPS
sensors. A.Danti [15] detected bump based on texture variation, Hough transform and local
neighborhood information. B.R [13] proposed Wolverine - a non-intrusive method that uses
sensors present on smart phones. M.A [16] describes a mobile sensing system for road
irregularity detection using Android OS based smart-phones. K.Ganesan [17] detects pothole and
speed bump using image processing, but they mainly focused on shadow cancelling using
OpenCV. VP.Tonde [18] describes about a mobile sensing system for road irregularity detection
using Android OS based smart phone sensors.
Most of the work mentioned above on speed bump detection are done with sensors,
accelerometer and smart phone applications. The information about speed bump are collected
and stored in the database initially and these offline data are used for testing the system. Thus it
cannot be suitable for real-time scenarios whereas, the proposed system in W.Devapriya [11] is
very much compatible with the real-time implementation. This methodology is applicable for the
standalone cars without disturbing the network and can be embedded in higher end vehicles,
especially self-driving cars. The average performance of the system considering only speed
bump with proper marking is 85%. This methodology does not need any external hardware,
sensors and smart phones. Without including the GPS network and disturbing the mobile battery,

detection of speed bumps can be made possible easily. Improvements to the existing system can
be done by concentrating on detecting speed bump during night time, detecting speed bumps
which do not follow any pattern or marking, training the speed bump detection system using
neural network concept and distinguishing zebra crossing with the speed bumps.
2.4 Steer by Wire System
Automotive Steering has evolved a lot from its mechanical gear driven system to completely
electronic system. The difference between these two are mentioned in Figure 13 below.
FIGURE 13: Difference between Conventional and Steer-by-wire steering system [19].
The steering wheel is connected to an angle sensor which detects the angle of rotation of the
wheel with respect to its default position. This data is then sent to the ECU (Electronic Control
Unit) as CAN (Controller Area Network) Signals. A description of CAN Communication protocol is
mentioned in the following section. The electronic control unit uses this steering angle and then
actuates a motor connected to the EPS module or the Electronic Power Steering module. The
angle of rotation of the EPS motor is correlated to the angle of the rotation of the steering wheel,
though they are not connected to each other through a steering column, but electronically,
through the ECU [19]. This mechanism is called Steer-by-wire or Drive-by-wire system.
Security for automotive embedded systems has been studied initially. S.Checkoway [20]
analyzed the external attack surface of a modern automobile. The authors discovered that remote
exploitation is possible via a broad range of attack vectors such as mechanics tools, CD players,
bluetooth, and cellular radio. M.L.Chavez [21] incorporated confidentiality service in CAN based
on RC4. However, the authors did not consider fault-tolerance (FT) aspects or CAN message
priorities while analyzing the security over CAN. Several earlier works explored dependability for
automotive embedded systems. M.Baleani [22] discussed various FT architectures for automotive
applications including lock-step dual processor architecture, loosely-synchronized dual processor
architecture, and triple modular redundant architecture. Although, the authors compared various
FT architectures’ costs based on the area estimates, the study did not quantify the architectures’
FT capabilities subject to real-time constraints. H.Schweppe [23] studied an active braking
system assisted by vehicle-2-X communication. C.Wilwert [24] presented an approach for
evaluating the temporal performance and behavioral reliability of an SBW system considering the
delay variation introduced by network transmission errors. The authors evaluated the maximum
tolerable system response time and the impact of this response time on QoS for a time division
multiple access communication protocol. However, the work considered only communication
transmission errors and not the computational errors in ECUs. Furthermore, the authors did not
consider the system vulnerability to security attacks and did not analyze the scalability aspects. In
A.Munir [19], the system considers security overhead while analyzing the system response time
and analyzes the SBW system’s response time under different load conditions and message
priority assignments.

The advantage of this system is a closed loop control as well as fault detection and analysis, as
steering is a highly safety critical mechanism in a car. Steer by wire also come handy in
developing autonomous cars as the ECU can now control the Steering by tapping in to the CAN
Bus and sending and receiving relevant angle info and motor turn. An SBW system eliminates the
risk of steering column entering the cockpit in the event of a frontal crash and thus enhances
safety. Since steering column is one of the heaviest components in the vehicle, removing the
steering column reduces the weight of the vehicle and therefore lessens fuel consumption. Steer
by wire systems enhance driver comfort by providing a variable steering ratio.
2.5 Controller Area Network
CAN is a serial communication standard with 3 layers: (1) Physical layer, (2) Data-Link layer, and
(3) Application layer. These layers are mentioned in Figure 14 below. Each node is a ECU, for eg:
Electronic Power Steering ECU, Body Control ECU etc.
FIGURE 14: A Typical CAN Bus Nodes in a Car [25].
All the ECUs talk to each other through the CAN Communication protocol in the vehicle. A typical
CAN message will have 8 bytes of data. Each byte can be allotted to specific info as per the
developers wish. Each message has an ID assigned to it. All nodes can send and receive data.
Also, CAN is a broadcasting protocol. So, it is not point to point, thus multiple ECUs can receive
info from a single sender. In the steer by wire e.g., Node 1 is the Power Steering ECU, node two
can be Vehicle Control ECU. The Power steering ECU sends Steering angle as 2 bits with
message ID 0x180. The receiving ECU listens for his message ID, as soon as it receives it, it
decodes the 2-bit info into steering angle and then activate the power steering motor [25]. The
same information can be received from Instrument cluster ECU and can display angle info in the
cluster and similarly the Body control unit can activate the turn signal automatically. Thus, it is a
highly scalable network, used widely in automotive field for decades.
The important parameters like Bus Load, Quality of Service of the network etc. are important in
analyzing the robustness of the network. As threats of remote hacking of cars have been made
recently, there is great emphasize of increasing the efficiency of CAN network.
2.6 NVIDIA Real Time Hardware Solutions for Autonomous Cars
NVIDIA DRIVE™ PX 2 is the open AI car computing platform that enables automakers and their
tier 1 suppliers to accelerate production of automated and autonomous vehicles. It scales from a
palm-sized, energy efficient module for Auto-Cruise capabilities, to a powerful AI supercomputer

capable of autonomous driving. The platform uses two Parker processors and two Pascal
architecture-based GPUs to power deep learning applications. Parker delivers up to 1.5 teraflops
(a unit of computing speed equal to one million-million (10^12) floating-point operations per
second) of performance for deep learning-based self-driving AI cockpit systems. It also has dual
CAN (Controller Area Network) channels, an automotive communication protocol standard. Its
main brain is four 64-bit ARM Cortex A57 CPUs.
The researchers trained a convolutional neural network (CNN) to map raw pixels from a single
front-facing camera directly to steering commands and this end-to-end approach proved
surprisingly powerful. With minimum training data from humans the system learns to drive in
traffic on local roads with or without lane markings and on highways. It also operates in areas with
unclear visual guidance such as in parking lots and on unpaved roads. The system automatically
learns internal representations of the necessary processing steps such as detecting useful road
features with only the human steering angle as the training signal without providing any explicit
training [26]. Compared to explicit decomposition of the problem, such as lane marking detection,
path planning, and control, this end-to-end system optimizes all processing steps simultaneously
and will eventually lead to better performance and smaller systems. Better performance will result
because the internal components self-optimize to maximize overall system performance, instead
of optimizing human-selected intermediate criteria.
This hardware setup can run a machine learning algorithm, which can take in real time of camera
images and steering angle from a manual run car and process those data to create an algorithm,
and then use the algorithm later to drive by itself. The three simple steps are mentioned with block
diagrams below [26]. Most of the algorithms mentioned above comes in Step 2 below.
Step 1: Data collection
Figure 15 shows a simplified block diagram of the collection system for training data. Three
cameras are mounted behind the windshield of the data-acquisition car. Time-stamped video
from the cameras is captured simultaneously with the steering angle applied by the human driver.
Training with data from only the human driver is not sufficient. The network must learn how to
recover from mistakes [26]. Otherwise the car will slowly drift off the road. The training data is
therefore augmented with additional images that show the car in different shifts from the center of
the lane and rotations from the direction of the road.
FIGURE 15: Data Collection using NVIDIA Drive PX in a manual car [26].
Step 2: Training the Neural Network
A block diagram of the training system is shown in Figure 16. Images are fed into a CNN which
then computes a proposed steering command. The proposed command is compared to the

desired command for that image and the weights of the CNN are adjusted to bring the CNN
output closer to the desired output. The weight adjustment is accomplished using back
propagation.
FIGURE 16: Training the NVIDIA Drive PX Artificial Neural Network Algorithm using the collected data [26].
Step 3: Autonomous Drive
Once trained, the network can generate steering from the video images of a single center
camera. This configuration is shown in Figure 17.
FIGURE 17: DRIVE PX driving autonomously by analyzing real time data using the trained Algorithm [26].
Training data was collected by driving on a wide variety of roads and in a diverse set of lighting
and weather conditions. Most road data were collected in central New Jersey, although highway
data was also collected from Illinois, Michigan, Pennsylvania, and New York. Other road types
include two-lane roads (with and without lane markings), residential roads with parked cars,
tunnels, and unpaved roads. Data was collected in clear, cloudy, foggy, snowy, and rainy
weather, both day and night. In some instances, the sun was low in the sky, resulting in glare
reflecting from the road surface and scattering from the windshield. The system has no
dependencies on any vehicle make or model [26]. Drivers were encouraged to maintain full
attentiveness, but otherwise drive as they usually do.
Evaluation of networks is done in two steps, first in simulation, and then in on-road tests. In
simulation, the networks provide steering commands in their simulator to an ensemble of
prerecorded test routes that correspond to about a total of three hours and 100 miles of driving in
Monmouth County, NJ. The test data was taken in diverse lighting and weather conditions and
includes highways, local roads, and residential streets. In simulation, what percentage of the time
the network could drive the car (autonomy) is estimated. The metric is determined by counting
simulated human interventions [26]. These interventions occur when the simulated vehicle
departs from the center line by more than one meter, under the assumption that, in real life, an
actual intervention would require a total of six seconds: this is the time required for a human to
retake control of the vehicle, re-center it, and then restart the self-steering mode. Equation (3)
represents formula for calculating percentage autonomy.
autonomy = (1 – (number of interventions).6 seconds).100 (3)
elapsed time [in seconds]

After a trained network, has demonstrated good performance in the simulator, the network is
loaded on the DRIVETM PX in the test car and taken out for a road test. For these tests,
performance is measured as the fraction of time during which the car performs autonomous
steering. This time excludes lane changes and turns from one road to another [26]. For a typical
drive in Monmouth County NJ from Holmdel to Atlantic Highlands, the system was autonomous
approximately 98% of the time.
3. FUTURE SCOPE
X.Li et al [1] experimentally showed that Spiking Neural network based edge detection is more
efficient than Canny edge detection method in the case of lane marking identification on roads.
The future work in this field would be identifying much more robust edge detection algorithms with
less computational time. A.Salhi [2] discusses the method of traffic sign identification using
Polygonal Approximation of digital curves technic applied on contours. The results are also
presented where real time traffic signs are classified and identified using this method. The future
work in this field would be analyzing other algorithms in traffic sign detection and figure out the
most optimum and efficient methods. Also, the samples traffic signs are limited in this paper. The
system should be able to scale the samples to include all the diverse varieties of traffic signs and
identify it precisely. W.Devapriya [4] discusses Speed bump detection using Horizontal projection.
The test results show that it could identify speed bumps with zebra crossing using this method.
The future work in this field would be identify all varieties of speed bumps with different markings
and no-markings. A.Munir [5] talks about Steer by wire systems and its security features. It also
exhaustively talks about Controller Area Network and how it is impacted when its network
parameters like Quality of Service Bus Load varies. The paper also defines fault tolerance,
multithreading and error detection of the CAN Bus network in a modern car. M.Bojarski [7] shows
the industrial case study done by NVIDIA to develop a self-driving car based on real time
embedded systems. It shows the systematic approach to be taken while using cameras as
sensors and Artificial Neural Network to teach the ECU to learn from the camera images and the
steering angle information from the controller. It is an exciting field of conglomeration of all the
latest technologies in automotive, image processing, computer vision, neural networks and
embedded systems.
4. CONCLUSION
Various image processing techniques have been presented through these papers which have high
end real time applications in the day to day life. The important take away terms are Computer
vision techniques, Noise reduction and Content extraction. Various algorithms and filters are being
used to achieve high efficiency data extraction from images. After evaluating the end results of the
papers analyzed, it can be concluded that, out of the several algorithms reviewed in the literature
survey, 75% were found success in real time in embedded systems. Real time environment
functionalities push the realm of embedded controllers to be used to execute these features.
Software development environment OpenCV has also been discussed in these papers. Also, there
are small abstract of how steer by wire system and CAN communication protocol is involved in an
automotive application. In the end a case study done by NVIDIA to develop an autonomous car
using these mentioned technologies is discussed.
5. REFERENCES
[1] X.Li, Q.Wu, Y.Kou, L.Hou, H.Yang, “Lane Detection Based on Spiking Neural Network and
Hough Transform” 2015 8th International Congress on Image and Signal Processing (CISP
2015)
[2] Y. Yuan, G. L. Zhang, “Improved Median Filtering Method,” Journal of Chengdu University of
Technology,Vol. 31, pp.126-129,2013
[3] Canny J, “A Computational Approach To Edge Detection”, IEEE Trans. Pattern Analysis and
Machine Intelligence, 8(6):679–698, 1986

[4] Q. Wang, “An Improved Lane Detection Algorithm Based on Hough Transform,” Computer
and Digital Engineering,Vol. 42, pp.2164- 2167,2014.
[5] Z. L. Jin, G. M. Hu, H. J. Zhang, G. H. Man, “Lane Detection Method Based on Optimal
Threshold and Hough Transform,” Instrumentation and Measurement,Vol. 28, pp.88-
91,2009.
[6] J. Guo, Z. H. Wei, D. Q. Miao, “Lane Detection Method Based on Improved RANSAC
Algorithm,” 2015 IEEE Twelfth International Symposium on Autonomous Decentralized
Systems,pp.285-288,2015
[7] J. R. Cai, J. W. Zhao, “Camera Calibration of Binocular Stereo Vision System,” Journal of
Jiangsu University,Vol. 27, pp.6-9,2006.
[8] A. Salhi, B. Minaoui and M. Fakir, "Robust Automatic Traffic Signs Detection using fast
polygonal approximation of digital curves”, Multimedia Computing and Systems (ICMCS),
2014 International Conference on, Marrakech, 2014, pp. 433-437.
[9] Urs Ramer, "An iterative procedure for the polygonal approximation of plane curves",
Computer Graphics and Image Processing, 1(3), 244-256 (1972).
[10] Dataset "The German Traffic Sign Detection Benchmark", http://benchmark.ini.rub.del?
section=gtsdb&subsection=dataset 2013.
[11] W. Devapriya, C. N. K. Babu and T. Srihari, "Advance Driver Assistance System (ADAS) -
Speed bump detection," 2015 IEEE International Conference on Computational Intelligence
and Computing Research (ICCIC), Madurai, 2015, pp. 1-6.
[12] M.Jain, A.P.Singh, S.Bali, S.Kaul, “Speed-Breaker EarlyWarning System”.
[13] B.R, N.Vankadhara, B.Raman and P.Kulkarni, “ Traffic and road condition estimation using
smartphone sensors” in Communication Systems and Networks (COMSNETS), 2012 Fourth
International Conference on (jan. 2012), pp. 1 –6.
[14] M.P, P.V. and R.R.; Nericell: rich monitoring of road and traffic conditions using mobile
smartphones. In: 6th ACM conference on Embedded network sensor systems, SenSys ’08,
pp 323– 336. ACM, New York, NY, USA (2008).
[15] Dr. A.Danti, J.Y. Kulkarni, Dr. P.S.Hiremath , “A Technique for Bump Detection in Indian
Road Images Using Color Segmentation and Knowledge Base Object Detection”,
International Journal of Scientific & Engineering Research, Volume 4, Issue 8, August 2013
ISSN 2229-5518.
[16] M.A, S.G, Z.R, K.G, And S.L, “ Real time pothole detection using android smartphones with
accelerometers” Distributed Computing in Sensor Systems and Workshops, International
Conference on (2011), 1–6.
[17] K.Ganesan, R.Natesan, A.Raj, N.Sekar , “An Image Processing Approach to Detect
Obstacles on Road,” SAE Technical Paper 2015.
[18] V. P. Tonde, A.Jadhav, S.Shinde, A.Dhoka, S.Bablade “ Road Quality and Ghats
Complexity analysis using Android sensors” International Journal of Advanced Research in
Computer and Communication Engineering Vol. 4, Issue 3, March 2015.
[19] A. Munir and F. Koushanfar, "Design and performance analysis of secure and dependable
cybercars: A steer-by-wire case study," 2016 13th IEEE Annual Consumer Communications
& Networking Conference (CCNC), Las Vegas, NV, 2016, pp. 1066-1073.

[20] S. Checkoway, D. McCoy, B. Kantor, D. Anderson, H. Shacham, S. Savage, K. Koscher, A.
Czeskis, F. Roesner, and T. Kohno, “Comprehensive Experimental Analyses of Automotive
Attack Surfaces,” in Proc. of the 20th USENIX conference on Security (SEC), August 2011.
[21] M. L. Chavez, C. H. Rosete, and F. R. Henr´ıquez, “Achieving Confidentiality Security
Service for CAN,” in Proc. of IEEE International Conference on Electronics,
Communications, and Computers (CONIELECOMP), March 2005.
[22] M. Baleani, A. Ferrari, L. Mangeruca, A. Sangiovanni-Vincentelli, M. Peri, and S. Pezzini,
“Fault-Tolerant Platforms for Automotive Safety-Critical Applications,” in Proc. of ACM
CASES, October-November 2003.
[23] H. Schweppe, T. Gendrullis, M. S. Idress, Y. Roudier, B. Weyl, and M. Wolf, “Securing
Car2X Applications with Effective Hardware- Software Co-Design for Vehicular On-Board
Networks,” in Proc. Of Joint VDI/VW Automotive Security Conference, Berlin, Germany,
October 2011.
[24] C. Wilwert, Y.-Q. Song, F. Simonot-Lion, Loria-Trio, and T. Cl´ement, “Evaluating Quality of
Service and Behavioral Reliability of Steer-by-Wire Systems,” in Proc. of IEEE ETFA,2003.
[25] A Rezaei, “An introduction to CAN (HEV Enterprise)”, CAN Tutorials.
[26] M Bojarski “End to End Learning for Self-Driving Cars” https://arxiv.org › cs, 2016.

Applications of Image Processing and Real-Time embedded Systems in Autonomous Cars: A Short Review

More Related Content

What's hot (20)

Similar to Applications of Image Processing and Real-Time embedded Systems in Autonomous Cars: A Short Review (20)

Recently uploaded (20)

Applications of Image Processing and Real-Time embedded Systems in Autonomous Cars: A Short Review