Lane and Obstacle Detection
Based
on
Fast Inverse Perspective Mapping Algorithm
Gang Yi Jiang, Tae Young Choi, Suk Kyo Hong, Jae Wook Bae*, and Byung Suk Song*
Ajou
University, Korea
*
Institute for Advanced Engineering, Korea
Abstract
A fast inverse perspective mapping algorithm (FIPMA) is
presented for a fast and accurate recovering of road surface
from a given 2D road image. FIPMA is able to simplify the
system design without losing reliability and flexibility. A
novel lane and obstacle detection method, using only a
single CCD camera, is proposed based on recovered road
surface image by FIPMA. This method includes five parts:
recovering road surface from input road image using
FIPMA, processing the recovered surface image, detecting
lane and estimating lane parameters, updating camera
parameters adaptively, and detecting obstacles.
1
Introduction
The growing volume of traffics in Korea requires higher
and higher level of traffic safety. Vehicle guidance system
in intelligent transportation systems uses image processing
andor machine vision to detect lane markings, vehicles,
pedestrians, road signs, traffic conditions, traffic incidents,
and even driver drowsiness [1,2]. Vision based vehicle
guidance system is helpful for relieving the contradiction
between traffic safety and traffic density.
In order to follow road and keep a vehicle on the right
lane and a safe distance from a front vehicle, lane and
obstacle should be first of all found from images recorded
by cameras mounted on the vehicle. Many techniques of
lane and obstacle detection have been developed such as
neural network based approaches
[3],
optical flow
techniques
[4],
identification of lane markings from color
information and deformable templates
[5],
model based
approaches
[6],
re-organized image based approaches
[7,8],
and
so
on.
Re-organized image based approaches use the inverse
perspective transformation to remove perspective effect and
recover road surface image. Recently, GOLD (Generic
Obstacle and Lane Detection) system
[7]
has been
developed. GOLD uses a stereo vision based hardware and
software architecture of three main parts; (1) detection of
road markings through morphological processing, (2)
alleviation of annoying problems caused by non-uniform
illumination, and
(3)
implementation of detection step on
massively parallel architectures to achieve real time
performance. But this system is quite vulnerable to road
conditions such as abrupt changing slopes. RALPH
(Rapidly Adapting Lateral Position Handler) system [SI
decomposes lane detection into three steps; (1) down-
sampling input image to create a low resolution image,
(2)
determination of road curvature, and
(3)
determination of
lateral offset of vehicle relative to the lane center. However,
RALPH has two problems. One is that the strategy of
down-sampling is effective to save large computational
expense of obtaining re-organized image, while the
resolution of down-sampled image is too low to extract lane
parameters effectively. The other is also on sloping roads.
In this paper, a new approach to lane and obstacle
detection is proposed. In section
2,
FIPMA is given in order
to
improve performance of system. In section
3,
a new
method of lane and obstacle detection
from
recovered road
surface image is described. Experimental results of lane and
obstacle detection with the proposed method are given in
section
4.
2
Recovery of Road Surface from a Road
Image with FIPMA
Fig. 1 shows a typical road image on highway, where the
road surface is divided into lanes by lane markings with
alternating dashed or solid lines. Lanes occupy quite large
area, especially current driving lane has larger area than its
neighbors. Lane boundaries are quite long and parallel to
each other. If a lane is straight in a 3D real world, it will
converge at a point in a 2D image plane, called a vanishing
0-7803-6583-6/00/$10.00
0
2000
IEEE
2969
资源评论
