IEEE Robotics & Automation Magazine - September 2017 - 66

Mobile Robots Localization
With the development of mobile robots, increasingly complex tasks are delegated to them. Vehicles are getting
smarter every day, and the applications are varied, such as
autonomous driving, systems that assist human drivers,
platooning [1], agricultural applications [2], assistance of
elderly people, and so on. To achieve those tasks properly, a
robot needs some awareness of its environment and reliable, accurate localization.
There are two kinds of localization methods, depending
on whether a map is already known. If a map is not known,
the robot has to be able to build the map itself to compute
its location. This is the well-known simultaneous localization and mapping (SLAM) method [3]. In the localization
step, the robot performs a bottom-up approach to deal with
a large amount of information that can be ambiguous,
noisy, or irrelevant [4].
The second approach is based on the assumption that the
robot performs its tasks in a well or partially known environment. This approach is very interesting for mobile robot
localization and navigation [5], as there is an increasing number of very complete maps (e.g., openStreetMap,
googleMap, etc.) available. The main advantage of such
methods is that the localization is world referenced and can
therefore be fused with absolute localization sensors like
GPS. It is possible to consider the information exchange
among vehicles without bias or reference problems. The presented method relies on an absolute localization and has
been partially described in [6].
Absolute Localization
The most popular absolute localization systems use global
navigation satellite systems (GNSSs): GPS, GLONASS
(Germany), and BeiDou (China). Today, GPS is used everywhere by everybody for driving, military applications, and
more. This solution is good for outdoor applications because
it is an easy and, generally, inexpensive solution.
However, GNSSs have some drawbacks: accuracy (generally between 1 and 10 m for common ones), a multipaths
problem, especially in urban areas [7], and it is subject to
interference, occlusions, clock errors, or signal deformations.
The problem of accuracy can be solved by methods like realtime kinematic and differential GPSs, which are very accurate but expensive and require equipping the environment
with ground-based reference stations. Moreover, availability
problems, multipaths, and occlusions still remain. To solve
these problems and obtain an accurate vehicle localization,
most works use data fusion, trying to improve GNSSs accuracy by using additional sensors. If a map exists, these sensors allow computation of the absolute position by retrieving
the features of the map.
A map [8] can be very simple, such as a topological map
for common driver navigation systems. An environment
map with roads [9] can also be used like a data source. The
methods coupling GPS and maps can provide very good
results but are generally restrained to navigation on a road
66

IEEE ROBOTICS & AUTOMATION MAGAZINE

SEPTEMBER 2017

and are not suitable for an accurate localization for automatic
vehicle guidance.
Other systems use extended maps, i.e, with landmarks
detectable with a specific sensor (e.g., lidar [10] or a camera
[11]). The challenge now is to fit the data provided by sensors to the map. This is the well-known map-matching problem [12]. But how can the information provided by different
sensors be consolidated? The localization problem is more
often solved in a particular environment, such as an indoor
environment [13] or urban area [14]. This discrimination is
based on the characteristics of both the map and sensors.
However, having different sensors and using advantages
from each one can significantly improve the performance of
the localization system.
An interesting method presented in [15] takes into account
the intrinsic differences between different sensors and is able
to use a sensor only if it is relevant. This approach is not bottom-up like the most popular methods in which the data from
each sensor are analyzed regardless of their relevance at a
given time. However, this top-down method, inspired by cognitive research, requires learning a specific environment and
works in replaying a previous learned trajectory.
These sensor management methods [16] are not yet used
much in vehicle localization. Overall, the principle of these
approaches is to select, at each instant, the most appropriate
sensor according to a criterion. The principle is to select
the best sensor by directly taking into account the goal of the
application and not the global information. The goal of sensor
management methods is to maximize the information content
of the posterior [17]. In the context of vehicle localization, this
can be seen by the accuracy of our estimation and its integrity.
This is the reason why we think that such sensor management can be a good solution to localize a mobile robot, as
well as optimizing both uncertainty and integrity. So we look
for a generic localization system that is able to:
● use a georeferenced semantic map that naturally allows a
collaborative localization between several vehicles
● manage different kinds of sensors, depending on the
context (e.g., indoor/outdoor)
● provide in real time (for robotic purposes) an accurate and
reliable estimation of the localization.
Top-Down Approach Proposed
To address the previous problems, we propose in this article a
new localization method in which every resource (sensor and
detector) is used according to a global objective criterion and
the current localization state. Figure 1 illustrates this process
in a top-down manner. The main four steps of the process
are presented.
1) Check if objectives are reached.
2) Select the best action.
3) Perceive the environment.
4) Update the state by taking into account the result of the
perception process.
This article is organized around the presentation of each step
with its description, including its interaction with the others.

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - September 2017