Abstract—Accurate localization and mapping play a pivotal
role in mobile robot navigation. In this paper, we present a novel
algorithm for mobile robot SLAM (Simultaneous Localization
and Mapping) based on stereo vision. First a novel method is
proposed to extract distinctive invariant image features, which
is coined PLOT (Polynomial Local Orientation Tensor). The
stability of these features to image translation, scaling, rotation
and illumination changes makes them suitable landmarks for
mobile robot visual SLAM. The visual landmarks relative to the
robot can be established by matching the PLOT features.
Mobile robot localization is achieved by matching these
distinctive landmarks in the current frame to the database map.
RANSAC algorithm is improved, coined extended RANSAC,
for mobile robot pose estimate due to its efficiency. Meanwhile,
the visual landmarks in the database map are updated
correspondingly. Experimental results show that the proposed
method based on the PLOT features achieves SLAM for mobile
robot with higher precision.
I. INTRODUCTION
N mobile robot applications, SLAM is a fundamental and
important ability to localize itself accurately and build a
map of the environment. It is a challenging research topic in
mobile robot which has received much attention over the last
few years [1-3]. Most of the existing SLAM algorithms are
based on sonar sensors, laser range finders or visions [4]. In
[5], the authors developed an algorithm for model-based
localization in a known environment that relied on the
concept of a geometric beacon observed in sonar
measurements. The algorithm was based on extended Kalman
filter. Triebel and Burgard constructed highly accurate 3D
models of the environment based on the edge features
extracted from a single 3D laser scan using the global
constraints [3]. Davison and Murray proposed an approach
for simultaneous localization and map-building using active
vision [6]. By comparison, vision systems are much more
informative and of high resolution. Moreover, the 3D
coordinates of the points from the scene can be obtained
while using stereo cameras. Therefore, vision-based
approaches using stable visual landmarks in the unmodified
environments are highly desirable for SLAM, denoted as
visual SLAM.
The distinctive image features [7] used in visual SLAM are
Manuscript received September 12, 2013. This work is supported by the
National Natural Science Foundation of China (61203332).
Rui Lin is with the School of Mechanical and Electric Engineering,
Soochow University, PR China. (e-mail: linrui@suda.edu.cn).
Maohai Li is with the School of Mechanical and Electric Engineering,
Soochow University, PR China. (
*
corresponding author, Tel:
+86-0512-67587229, e-mail: limaohai@163.com).
Lining Sun is with the School of Mechanical and Electric Engineering,
Soochow University, PR China. (e-mail: lnsun@hit.edu.cn).
extracted as visual landmarks, which depend on techniques
used like Harris [8], SURF [9], and SIFT [10]. Harris's 3D
vision system DROID [11] determined both the camera
motions and the 3D positions of the features by using Kalman
filter to track the Harris features. A vision-based mobile robot
localization and mapping algorithm was proposed based on
SIFT features with their Triclops stereo vision system [12]. In
[13], the local maps built by the robots consisted of the 3D
coordinates of the Harris points detected and their
correspondent descriptor U-SURF.
First local features should be robust to changes of viewing
conditions in order to allow for correspondences and also be
invariant to serious changes in color and intensity that may
exist between views. PLOT features presented in the paper
are suitable landmarks for mobile robot visual SLAM. Then
the features matching based on the nearest neighbor ratio [14]
is implemented to find correspondences for each frame
between the two different images of the same scene. After
finally matching the features, we would like to maintain a
database map containing the PLOT features extracted and to
use this database of landmarks to match features found in
subsequence views. The task of localization is to find the
robot pose in the global coordinate system. Extended
RANSAC is proposed for mobile robot accurate localization.
First RANSAC [15] algorithm is used to match the landmarks
obtained in the current frame with the pre-built visual
landmark database in the previous frames to compute mobile
robot pose. Then Hough Transform [16] and Least-Square
Minimization [17] are also used for an accurate pose estimate
and hence better global localization. The database of
landmarks is also incrementally updated over subsequent
frames and adaptive to the dynamic environment as mobile
robot moves in the scene.
II. E
XTRACTING STEREO IMAGE FEATURES
PLOT detector is proposed to extract scale and rotation
invariant features as the stable visual landmarks. It represents
the orientation tensor estimation based on the polynomial
expansion in detection of local image features. The features
are localized based on the small eigenvalues of the
normalized local orientation tensor by local maximum search
in the eight neighborhoods. Then each feature is assigned an
orientation and described by a descriptor based on the local
image region.
A. Detecting the Keypoints
Tensor is a useful tool in image analysis for the detection of
low-level features such as edges, lines, corners, and junctions.
It can describe local image properties in a way that is
Image Features-Based Mobile Robot Visual SLAM
Rui Lin, Maohai Li
*
, Lining Sun,
I
978-1-4799-2744-9/13/$31.00 ©2013 IEEE
Proceeding of the IEEE
International Conference on Robotics and Biomimetics (ROBIO)
Shenzhen, China, December 2013