![](https://csdnimg.cn/release/download_crawler_static/18591976/bg1.jpg)
ISSN 1054-6618, Pattern Recognition and Image Analysis, 2016, Vol. 26, No. 3, pp. 565–571. © Pleiades Publishing, Ltd., 2016.
Traversable Region Detection Based on Fusion-features
and Partial Least Squares
1
Hu Bin
School of Computer Science and Technology, Nantong University, Nantong, Jiangsu, China
e-mail:nj_chris@126.com
Abstract—In order to detect the traversable region of automotive land vehicle (ALV), a multi-scale data anal-
ysis and representation method, shearlet transform is researched. Based on the feature that the descriptor calls
Histograms of Shearlet Coefficients (HSC), a weighted HSC (WHSC) is proposed. Compared to HSC,
WHSC uses the scale factor, which makes it better than HSC. We combine WHSC and color histogram in
HSV color space as the fusion-feature, and use partial least squares (PLS) to project the high dimensional
feature vectors onto a subspace. Also, support vector machine (SVM) is used based on linear kernel as the
classification to solve traversable region detection. The experiment results suggest that for both in NUSTrobot
dataset and OUTEX dataset, the method provided by this paper performs much better, and can detect the tra-
versable regions in complex environments (e.g., different shadow and lighting conditions). Moreover, with
the help of this method, the platform can achieve more functions.
Keywords: traversable region detection, support vector machine, shearlet transform, PLS
DOI: 10.1134/S1054661816030068
1. INTRODUCTION
With the development of computer technology,
artificial intelligence technology, sensor technology as
well as robot theory and application have also being
developed rapidly. Among them, autonomous land
vehicle (ALV) is a kind of mobile platform which can
autonomously complete tasks of maneuvering, help-
ing or replacing human in specific combat situation
over complex terrains. Reliable traversable area detec-
tion is one of the most important technologies of ALV,
and it has been widely studied by institutions and
scholars.
The mainstream method of traversable area detec-
tion is based on machine learning. A classifier is
trained off-line, and while detection, sliding-window
method is used to calculate features of each window in
the picture. After the sliding-window scans the whole
picture, the detection finishes. Castano has used
Gabor and Gaussian Mixture Model (GMM) to
detect traversable area [1]; Shi has combined Gabor
and color feature using Bayes fusion method [2]; and
Kim has used Ultra-Pixel for traversable area detec-
tion [3]. In order to solve the problem of data markup,
Dima has used active learning for traversable area
detection [4]. Gao has used DCT and color feature as
the feature, and one-class SVM as the classifier to
detect traversable area [5].
1
The article is published in the original.
As for image, the information is mostly contained
in the edge and contour, and it cannot be well
described by the traditional one-dimensional wavelet
transform. Even two-dimensional wavelet decomposi-
tion is also confined to three directions, including hor-
izontal, vertical and diagonal. In order to overcome
the defects of traditional wavelet transform, lots of
image representation methods have been proposed in
recent years, such as Wedgelets transform [6], curvelet
transform [7] and contourlet transform [8], while cur-
velet transform is the most famous one among them.
But curvelet transform is not directly related with
multi-resolution theory, and it is difficult to be dis-
cretized. Guo and Easley have constructed a multi-
scale geometric image representation method, i.e.,
Shearlets [9], which describes image in multi-scale
and multi-direction. We improved the Histograms of
Shearlet Coefficients (HSC) [10] and combined the
improved feature and color feature as the fusion fea-
ture, and used partial least squares (PLS) for dimen-
sion reduction.
2. HISTOGRAMS OF SHEARLET
COEFFICIENTS
2.1. Histograms of Shearlet Coefficients
The shearlet and its decomposition method have
been proposed in [11], which provides a Parseval
framework for analyzing and representing data with
anisotropic information at multiple scales. Let f be an
APPLIED
PROBLEMS
Received March 15, 2015