Anda di halaman 1dari 5

A Robsut Neural Network Multi-Lane Recognition System

Amer K. Dawoud', Salah G. Foda* and Ahmad S. Tolba**

* Department of Electrical and Computer Engineering,


** Department of Physics,
Kuwait University, P.O. Box 5969, Safat 13060, Kuwait
E-mail: sfoda@eng.kuniv.edu.kw

AbsZruct: In this paper, the design and neural imp- vehicle. It should not show any tendency to
lementation of a vision based multi-lane highway break down when the environmental
lane recognition system are presented. The design variables such as road width, road conditions
objective of the system is to recognize the lane which and light conditions changes, and in the
a test vehicle is currently driving through by presence of external noise such as nearby
determining its left and right lane boundaries. When
passing vehicle.
the proposed lane recognition system was tested it
showed very high percentage of correct results in very
difficult circumstances which suggests that it
provides the basis for a reliable road following
system.
I. Introduction
For the past decade many institutes and research
centers have been conducting research on vision
based guidance systems for the purpose of
autonomously navigating land vehicles [1-91. The
building of such intelligent systems is a complex
and challenging task but with huge potentials in System low level medium high level
both civil and military application domains. level
SCARF r1-31 probability *** match filter
Autonomow land vehicle navigation system is a ALVINN [MI *** *** back prop.
system that can carry out tasks carried out by a ***
RBFN U] t**
radial basis
human driver. These tasks include road following,
speed control and tactical driving tasks such as Bristol [8,9] pixel voting *** match filter
obstacle avoidance, changing lane and negotiating
intersections. There are practical difficulties Table 1: Existing navigation systems:
related to the implementation of autonomous
navigation of an unmanned vehicle in the normal 11. Design of Lane Recognition System
roads. However, road following systems seem the
suitable and practical implementation where the The highway has been chosen as our
road following system can be used to monitor the environment in this work because it is less
vehicle's lateral position inside the road and to complicated than other road environments.
detect any unintended road departure. The components of road environment include
only asphalt road, lane markings, vehicles
We believe that any real progress towards and non-road surrounding areas like green
the implementation of vehicle navigation or plants or sand areas. It does not include
road following systems in normal roads intersections nor stop signals
should start with developing the recognition
system that could robustly produce correct The expected result from the proposed
judgments regarding the position of the recognition system is to identify the left and
right boundary lines of the lane that the 111. Road and Non-road Classification Network
vehicle is currently driving through. This part of the highway lane detection system
In the proposed lane recognition system, we represents the low level of processing where each
have chosen to follow the same steps of lane
pixel of the preprocessed color is classified into
recognition by human driver using neural
either road or non-road. The input space of the
networks. It relies on image processing at
preprocessed color image is of dimension 3, and
all the three levels. The lowest level of
each pixel is represented by a vector in a hyper
recognition process involves classifying the
cube of red green and blue components.
input image pixels into road and non-road
using Learning Vector Quantization
Building a robust vision system capable of
techniques. The medium level of recognition
recognizing highway lane in many different
process involves scanning the blocks of the
conditions requires the diversity of the
binary image resulting from the classification
highway image scenes used to derive the road
process searching for the lane markings.
and non-road color models required for the
Classical Back Propagation network was classificatior.. The diversity of these highway
trained for this purpose. Fig.1 shows the
scenes is assured by taking intc
modules of the proposed lane recognition
consideration the main factors that would
system. differentiate an image from another in a
highway environment. These main factors

+-I
input color image are:
(a) Color of the highway asphalt,
(b) Presence or absence of external noise of

P=+---l
re rocessin

[road and non-road ]


nearby passing vehicle,
(c) Shape of the white lane markings which
can be solid long line or dashed line,
(d) Nature of non-road area surrounding the
highway which can widely vary.
lane marking
Many highway video recordings were made
detection
v available and from these recordings we have
selected 250 images that represent a wide variety
lane boundary of highway scenes. Then each image is
determination preprocessed by light variation compensation,
then sub-samptng and averaging 2x2 block of
Fig.]: Block diagram for the proposed lane pixels into one pixel in the reduced image as
recognition system shown in the previous chapter. From each of these
images 4 pixels were selected to represent road
A SONY CCD video camera recorder is used input vectors. The total number of road input
to capture video recordings that are saved in vectors is 1000. The same was repeated for non-
the form of successive frames. The road areas. Rozd input vectors included asphalt
preprocessing module involves two road areas only, whiie the non-road input vectors
operations: Light variation compensation and included all elements of highway environment
Image size reduction. The i w g e reduction except the asphalt road such as:
replaces a block of pixels in the original (1) Highway surrounding areas like green plants
image with one pixel in the reduced image. areas and desert areas and the blue sky,
It was empiricalJy round that the best results (2) White highway lane markings,
for image reddction is obtained by using sub- (c) Barriers between the two sides of the highway,
sampling and averaging. (d) Obstacles like nearby passing vehicles of
different colors.
180
%wmLer 14 -16 1998

Using Matlab Neural Network Toolbox, learning IV. Highway Lane Detection Network
Vector Quantization GVQ) [lo-121 w a s utilized This part of lane recognition system represents
for the purpose of optimizing the location of the
the medium level of processing. The binary image
weight vectors. After training is completed, the resulting from classification stage is scanned
resulting LVQ network can be used to classify the searching for the highway lane markings. Medim
highway road images. level means that “blocks” of pixel areas are being
processed not single pixel as in the previous low
Each pixel of the preprocessed image is classified level classification stage.
into either road or non-road by presenting it to the
LVQ network of Fig. 2 and the output will be To locate the boundaries of a highway lane we
either [0 13 for road or [l 01 for non-road. The need first to find the positions of the white lane
size of binary output image resulting from this markings which were classified as non-road areas
stage is 80x60 pixels. inside road areas. These lane markings could be
defmed as:- white non-road areas surrounded by
Input Competitivelayer Linear layer Output road areas from both sides and have certain
n l shapes. The shape of the lane marking could be
also used as a feature that distinguishes it from
other non-road areas such as nearby passing
vehicles.

The factors that contribute in variation of the


shape and size of the white lane marking are:
(1) The type of the lane marking whether solid
white or dashed line,
Fig. 2: LVQ Classifying Network (2) The size of the lane marking changes with
change of the position of the marking with respect
to input camera. The closer the lane markings to
the camera the larger they will appear in the
image,
(3) The orientation of the lane rrxking with
respect to the input camera. Lane markings
located on the left of input camera will have
different orientation compared to lane markings
located on the right on input camera.
Fig. 3: Input image and the classification binary
output. Using a supervised neural network algorithm, the
proposed system was trained to detect various
types of lane markings. The inputs to the lane
marking detection system are blocks of the binary
pixel image resulting from classlfylng the
preprocessed image into road and non-road. The
output is a binary number 0 or 1 indicating the
absence or presence of lane markmgs in that
block. With this direct and defrnite input\output
relationship the best choice was to use back-
Fig. 4: Road classifcation for highway image
propagation Deural network algorithm with
which includes a dark vehicle
Levenberg-Marquardt approximation technique
for updating the weights and biases.
The first stage of the lane detection system is to Assuming that the trained feed foreword network
build models for the shapes of lane markings that could detect lane markings shapes of non-road
need to be recognized. areas surrounded by road areas, then to satisfy the
definition of lane markings we must make sure
that non-road areas are white colored.

In the previou:; section, the non-road color space


was divided into 6 clusters, three of them
represented light colors and the other three
represented dark colors. The white color of lane
marking of course belongs to the light colored
clusters. Which means that all non-road pixels in
5x2 window should belong to light colored
clusters. And if any of non-road pixels belonged
C
to dark colored ciuster then that window will be
Fig. 5(a): Original input image given output of 0 indicating the absence on lane
marking in that window.
(b): Binary classification
(c): 5x2 lane marking model
To detect white lane marking, a 5x2 window will
scan the binary image resulting from the road and
Fig. 5(a) shows an original color image, while
Fig. 5(b) shows the classification into road and non-road classification. This window will be the
non-road. Fig. 5(c) shows a lane marking model, input to the back-propagation network that will
a box of white non-road area surrounded by road highlight the positions in the image which match
area from both sides. Since close lane marking to any of the 42 models of lane marking. This
the input camera have a width of 3 pixels, while window will move with one pixel increment at a
far ones had a width of only one pixel.( 1 pixel = time.
2x2 pixels of original image), the width of the
The white color detection neural ndwork checks
block model was chosen to be 5 pixels so that it
can fit all types of lane marking in the midde of classification clusters of 5x2 window pixels and if
the model plus road pixels from left and right any non-road pixel does not bejong to light
sides. colored clusters, then the output of the trained
feed-foreword will be forced to be 0 indicating the
The architecture of the used back-propagation absence of liane marking in that window
network [13,14] is chosen as follows: regardless the output of trained feed-foreword
a) The number of inputs to the network = 10, network.
which represents the size of the 2x5 models in
Fig. 5(c). b) The number of neurons in the output
layer is one, indicating the presence or absence
lane marking.
c) The number of neurons in the hidden layer was
chosen to be 8. This number of hidden layer
binary
neuron was enough to reach the desired learning classification
accuracy. i nnge
d) Log sigmoidal activation functions in hidden
Road classifisd pixe!
and output layers neurons were chosen.
Non-rozd classified pixel

V. Design of Lane Marking Detection Network Detected lane marking

Lane markings could be defined as white non-road


areas surrounded by road areas from both sides. Fig. 6: Lane marking detection system.
References:
[ 11 V. Graefe and K.D. Kuhnert “Vision-based
autonomous road vehicles”, In Vision-based
vehicle guidance, Masaki I.(eds), Springer-Verlag,
NY,1992. .L

[2] J.D.Crisman and C.E. Thorpe, “SCARF: A


color vision system that tracks roads and
intersection”, IEEE Tram. on robotics and
automation, 1993.
[3] M. Meng and A.C. Kak, “Fast vision-guided
mobile robot navigation using neural networks”,
IEEE Proc. of system, Man and Cybernetic,
1992.
-
Fig. 7: Examples of original - and lane
images [4] D.A. Pemerleau, “Rapidly adapting artificial
marking detection output. neural networks for autonomous navigation”,
Advances in neural information processing,
VI. Conclusions and Future Work Morgan Kaufrnann Publiihers, 1991.
[5] T. Jochem and D. Pomerleau, “Life in the Fast
The developed lane recognition system Lane: The Evolution of an Adaptive Vehicle
performs a task that originally is carried out Control System”, AI wgazine, 1996.
by human driver. It can provide a skeleton [6] M. Rosenblum and L. Davis, “An Improved
for future work in two main directions Radial Basis Functiori Network for Visual
related to active areas of Artificial Autonomous Road Following”, IEEE Trans. on
Intelligence as follows: Neural Networks, 1996.
(a) real time implementation of lane [7] K. Dowling et al., “Navlab: An autonomous
recognition system which involves the navigation testbed” in Vision and Navigation:
interface with input color camera and The Carnegie Mellon Navlab, C . Thorpe (Ed.),
building dedicated hardware that satisfies Norwell, MA: Kluwer, 1990.
the real time requirements of processing each 181 F.W.J. Gibbs and B.T.Thomas, “The fusion
frame. of multiple image analysis algorithms for robot
road following”, IEE Fij3h Intemtional
(b) The development of road following or lane Conference on Image Processing and its
keeping system that utilizes the proposed fane Applications, 1995.
recognition system in finding the boundaries of [93 S. Spiegle, “Autonomous road vehicle
the current lane and then guide the vehicle to the navigation”, Proc. Electronic Technology to the
middle of the lane. The development of lane year 2000, 1995.
keeping system involves the generation of steering [lo] Y.Linde, A. Buzo, and R. M. Gray, “ An
angle decision, steering control mechanism and algorithm for vector quantizer design,” IEEE
the feedback from the vehicle steering mechanism Trans. Commun., 1980.
[113 J. Bryant, “On the clustering of multi-
The highest level of recognition process dimensional pictorid data,” Pattern Recognition,
involves selecting the markings which 1979.
belongs to the current lane and isolating [I21 T. Kohonen, “Self-organizing and
them from the total detected lane marking Associative Memory,” Spinger-Verlag, 1989.
set. This can be done using Fuzzy logic [13] M . H. Hasson, “Fundamentals of Artificial
which is the subject of current research. Neural Networks” The MlTpress, 1995.
Cl41 H. Demuth and M. Beale, ‘‘ Neural Network
TOOLBOX for use with Matlab,” The Math
Works Inc, 1994.

Anda mungkin juga menyukai