Abstract—This paper describes an efficient SLAM system
only using RGBD sensors. This system utilizes the Microsoft
Kinects to provide visual odometry estimation and 2D range
scans. The Kinect looking up toward the ceiling can track the
robot’s trajectory through visual odometry method, which can
provide more accurate motion estimation compared to wheel
motion measurement and cannot be disturbed under wheel
slippage. This is because the Kinect can provide a color image as
well as depth information such that robust 3D feature points
matching using invariant 2D feature descriptors such as SURF
and FAST is possible. Furthermore, the straight line features on
the ceiling can provide additional constraints on the inter-frame
motion of the camera and the loop closure leading to a more
accurate pose estimate. While the other two contiguous
horizontal Kinects can provide wide range scans, which ensure
more robust scan matching in the
RBPF-SLAM framework. In
addition, we develop a novel proposal distribution that relies on
visual odometry by replacing the transition motion model to
towards a SLAM solution. Subsequently, the accurate grid map
is online learnt through the adaptive resample
Rao-Blackwellized particle filter. Finally, our experimental
results, using three Kinects carried on mobile platform of
TurtleBot, clearly show the performance of our method.
I. INTRODUCTION
HE ability to learn a model of the environment and to
localize itself is one of the most important abilities of
truly autonomous robots able to operate within real world
environments. The problem of achieving this is one of the
most active areas in mobile robotics research, which is stated
as the simultaneous localization and mapping (SLAM)
problem.
The solution of the SLAM problem has been one of the
notable successes of the robotics community over the past
decade [1-7]. SLAM has been implemented in a number of
different domains from indoor robots to outdoor, underwater,
and airborne systems. There has been a wealth of research
into the SLAM problem in recent years, with reliably working
solutions for typical indoor scenarios using
Rao-Blackwellized particle filters like Gmapping being
available as open source software. However, these solutions
Manuscript received September 15, 2013. This work is supported by the
National Natural Science Foundation of China (61203332) and the Natural
Science Foundation of Jiangsu (SBK201321351).
Dr Maohai Li is with the School of Mechanical and Electric Engineering,
Soochow University, China, 215006 (e-mail: limaohai@163.com).
*Dr Rui Lin is with the School of Mechanical and Electric Engineering,
Soochow University, China, 215006 (corresponding author, phone:
0512-67587229; e-mail: linrui@suda.edu.cn).
Prof Han Wang is with the School of Electrical and Electric Engineering,
Nanyang Technological University, Singapore (e-mail: hw@ ntu.edu.cn).
Mr Hui Xu is with the School of Mechanical and Electric Engineering,
Soochow University, China, 215006 (e-mail: yanwu19@ 126.com).
work best in planar environments, rely on available,
sufficiently accurate odometry. For unstructured
environments, that lead to significant roll and pitch motion of
the carrier, such systems are not applicable or have to be
modified significantly.
Although there have been significant advances in developing
efficient SLAM algorithms for many different environments
at a theoretical and conceptual level. However, substantial
issues remain in practically realizing more general SLAM
solutions depending on different sensor technologies and
hardware platform, most methods are designed for highly
accurate laser range finders, which are much expensive.
There are few methods to consider the low-weight,
low-power and low-cost hardware for efficient general
SLAM implementation.
In this paper, we present a flexible and efficient system for
solving the SLAM problem that has successfully been used
on indoor environments. This system utilizes the Microsoft
Kinects to provide visual odometry estimation and 2D range
scans. The Kinect looking up toward the ceiling can track the
robot’s trajectory through visual odometry, which can
provide more accurate motion estimation compared to wheel
motion measurement. This is because the Kinect can provide
a color image as well as depth information such that robust
3D feature points matching using invariant 2D feature
descriptors such as SURF and FAST is possible.
Furthermore, the straight line features on the ceiling can
provide additional constraints on the inter-frame motion of
the camera and the loop closure leading to a more accurate
pose estimate. While the other two contiguous horizontal
Kinects can provide wide range scans, which ensure more
robust scan matching in the RBPF-SLAM framework. In
addition, we develop a novel proposal distribution that relies
on visual odometry replacing the transition motion model to
towards a SLAM solution. Subsequently, the accurate grid
map is online learnt through the adaptive resample
Rao-Blackwellized particle filter. The proposed system only
needs the low-weight, low-power and low-cost computational
resources which is modular, and can be easily portable for
different autonomous systems. Furthermore, this SLAM
system is implemented with the ROS operating system as
open source software. It honors the API of the ROS stacks
and thus can easily be interchanged with other application
such as localization and navigation stacks available in the
ROS system.
The main contribution of this paper is:
1. Development of a small-size of fully modular and portable
SLAM system that relies only on the low-weight, low-power
and low-cost computation resource and vision sensors for
online building maps of the indoor environment.
An efficient SLAM system only using RGBD sensors
Maohai Li, Rui Lin*, Han Wang, and Hui Xu
T
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