IEEE Robotics & Automation Magazine - December 2013 - 84

real-time visual SLAM with an extended Kalman filter (EKF).
This method shows an increase in localization accuracy by the
use of an explicit frame-to-frame error-propagation model, as
well as by the particle filter representation of feature depths for
initialization. In [18], this framework is extended by proposing
an inverse parameterization of feature depths. Since this allows
feature depths to be better approximated with a Gaussian, no
particle filter is required for initialization. However, the classical EKF-based SLAM gives rise to a fundamental limitation in
the number of features that can be mapped due to the quadratic complexity of the EKF algorithm. In [5], this technique
is integrated in a hierarchical SLAM framework based on the
aforementioned approaches and proposed as a means of building a large-scale environment map.
Pose Graph Optimization Technique
The most successful methods currently in use for solving
large SLAM problems with many loops are the pose graph
optimization algorithms [13], [20], [23]. Instead of estimating
the joint density of the robot pose and the landmark locations
observed, these methods find a set of poses that minimizes
the total error defined by the relative pose constraints, given
the recorded observations and control inputs. Such a scenario
is attractive because it is possible to transform the SLAM
problem to satisfy a generally sparse set of pose constraints
[20], [23], [24]. Like other methods, pose graph solvers have
the worst-case complexity at loop closure, which is dependent
on the number of poses to be estimated.
Limitation of Vision-Only SLAM
The main problem with most existing visual SLAM implementations that operate without odometry is their lack of
robustness in adverse situations. Tracking systems typically
rely on the prior pose used to limit the search for visual feature
correspondences, yielding a rapid frame-to-frame localization.
However, tracking failures occur frequently in situations where
vision is no longer reliable, e.g., under temporary sensor occlusions or featureless scenes like those caused by unlit spots in
the environment. For this reason, the incorporation of odometry into a visual SLAM is preferred for realistic applications.
Visual SLAM Using Odometry
Researchers who have used odometry and cameras have demonstrated reliable and accurate vision-based localization and
mapping [19], [21], [22]. In particular, schemes based on cameras pointing straight up at the ceiling [6], [11], [17] have certain advantages compared to other schemes in indoor environments. First, there is no significant effect caused by
dynamic obstacles such as moving people since the camera
points at the ceiling. Additionally, there is no scale change of
features between neighboring images in most indoor environments. Therefore, under the aforementioned conditions, visual
features can be extracted quickly from an image while robust
matching results are attained. In [6], a camera pointed toward
the ceiling is used to match the current image to a large ceiling
mosaic covering the whole operational space of the robot. The
84

IEEE ROBOTICS & AUTOMATION MAGAZINE

DECEMBER 2013

mosaic has to be constructed in advance, which involves a
complex state estimation problem. Furthermore, it is not
designed for a low-cost embedded system. In [17], a visionbased SLAM is proposed for low-cost applications, but it still
lacks the quadratic complexity of the EKF algorithm.
Motivation, Contribution, and Organization
There are very few works with a lightweight and low-cost
approach targeting commercial applications such as home
cleaning robots. Most general-purpose visual SLAM algorithms are not appropriate for low-cost applications because
they require high-resolution images or rapid processing
power. Therefore, it is worthwhile to investigate a visual
SLAM application for lightweight, low-cost applications. To
this end, a simple low-resolution camera with an ARM11
processor may be a good choice for implementing such an
embedded visual SLAM.
In this article, we present an efficient visual SLAM algorithm for resource-limited consumer robots such as home
cleaning robots. The proposed SLAM algorithm uses odometry and a single camera with a wide-angle lens that yields
low-resolution images of 320 # 240 pixels. The camera
points in the zenith direction perpendicular to the ground
plane and is restricted to approximate only two-dimensional
(2-D) planar motion. Under these conditions, the upwardlooking camera dramatically reduces the computational
complexity needed for extraction of features and matching
compared to the front-view cameras. The original contributions of this article are:
● a visual algorithmic compass that accurately estimates the
robot's orientation using the orthogonal lines of an environment detected in upward-looking scenes
● computing the three-dimensional (3-D) locations of features and the loop closure constraint using an upwardlooking camera
● reducing the linearization errors and the computational
complexity in the pose graph optimization technique
described in [23] using the visual algorithmic compass
● integrating a standard Kalman filter-based localization
method into the SLAM framework for efficiency
● development of a lightweight and robust SLAM algorithm on an embedded system for practical applications.
The article from here is organized into the following sections. An overview of the proposed SLAM algorithm is
described in the "Algorithm Overview" section. Methods for
defining spatial constraints based on the upward-looking
camera and correcting the robot trajectory using the spatial
constraints are presented in the "Visual Slam Framework"
section. A number of experimental examples to demonstrate
the validity of the proposed method are given in the
"Experimental Results" section, while the results are detailed
in the "Conclusions" section.
Algorithm Overview
An overview of the proposed visual SLAM algorithm consisting of six main modules is shown in Figure 1. The six main

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - December 2013