Error Analysis of Fisheye Correction Curve
Gang Bi, Xiaoling Zhang Weijia Feng, Junchao Zhu, Xinya Lv
Tianjin Key Laboratory for Control
Theory and Applications in Complicated Systems
School of Electrical Engineer
Tianjin University of Technology
391, Binshui Xidao Xiqing District, Tianjin, 300384, China
Tianjin University of Technology
391, Binshui Xidao Xiqing District, Tianjin, 300384, China
13324486335@163.com shanxi_zxl@sohu.com
Abstract–Various types of image can be captured with
fisheye lens, their wide field of view is particularly suited to a
stereo vision. However, fisheye lens introduces distortion, this
change makes the image difficult to identify. If a correction curve
can be fitted to the distortion shape, the degree of distortion can
be described, support vector machine is a popular machine
learning method for classification and regression, support vector
classification can find optimal interval of the different classified
data, the classified curve is as the correction curve, the training
data is obtained by corner detection, the center of the data is
gained by Hough transform. Although ignoring the error in the
process of the algorithm, but the error is still existed during the
picture taken from a fisheye lens in different situation, for
example angle or distance, then using the support vector
regression to gain the error between the original data and
predicted data, original data stem from training target which is
also obtained by a fisheye lens and predicted data stem from
another fisheye image.
Key words: fisheye, distortion, correction, classification,
regression, support vector machine
I. INTRODUCTION
At present, stereo vision technology are popular fields of
visual processing[1], we can get image information in a wide
field of view by a panoramic vision system and we can get
calculate the information of the object with a stereo vision
system[2]. A fisheye lens can have many advantages over
traditional lens, when used for such things to realize a stereo
vision. It is able to see half of the world. It can give a camera a
complete view of
180
D
along the horizon and
180
D
vertically
when pointed forward. However use of the fisheye lens brings
with its own problems, the image taken from a fisheye lens is
distorted, the distortion renders them hard to understand and
uncomfortable to watch. The Fig.11 shows what a picture
taken from a fisheye lens look like. In order to establish the
correction curve for the fisheye lens camera, the intrinsic
parameters must be determined. About the parameter of the
center of the training target, the optical center of any lens is
defined as the point where the optical axis passing through the
lens intersects the center of training target, using the Hough
transform algorithm to obtain the coordinates of center. To
obtain quantitative measurements from fisheye lens images,
we can use the corner detection to calibrate certain target
parameters coordinates for the training data in support vector
machine (SVM) algorithm.
Support vector machine were already known as a tool
that discovers information, SVM are also very effective for
classification among different features, a vector that we call a
pattern of
n components which we call feature, the features
are corner coordinates and the patterns are correspond to
labels which is different distorted area, we limit ourselves to
two-class classification problems. The training patterns are
used to build a decision function which is our correction
curve, new features are classified according to the sign of the
decision function. A problem in classification specifically, and
machine learning in general, is to find ways to reduce the
dimensionality of the feature space to overcome the risk of
overfitting, data overfitting arises when the number of features
is large, training techniques that use regularization avoid
overfitting of the data without requiring space dimensionality
reduction, projecting on the few principal directions of the
training data is a method commonly used to reduce feature
space dimensionality, with such a method, new features are
obtained that are nonlinear combinations of the original
features, eliminating some of original input features and
retaining a minimum subset of features yield classification
performance, the quality and complexity of the support vector
machine solution does not depend directly on the
dimensionality of the input space[3], based on these
advantages, we used support vector machine for data
classifying and fitting.
II.
MODELING OF THE AMPHIBIOUS ROBOT
A. Fisheye image pixel storage
Fig.1 The image of the pixel shortage
As shown in Fig.11, straight lines are bended under the
fish-eye cameras. With the corresponding pixel distribution,
there has been a distortion in Fig.1, according to the principle
of fisheye lens image, the degree of distortion is not the same,
1659
978-1-4799-7098-8/15/$31.00 ©2015 IEEE
Proceedings of 2015 IEEE
International Conference on Mechatronics and Automation
August 2 - 5, Beijing, China