Page 177 - Autonomous Mobile Robots

Landmarks and Triangulation in Navigation 161

bymatchingtheSIFT(scaleinvariantfeaturetransform)features. Feature-based
methods are often very efficient, and we have adopted it in our design.
However, the presence of nonunique feature landmarks causes the serious
concern in feature-based visual navigation. Therefore, instead of undistinguish-
able landmarks addressed in the previous section, we propose a new type of
artificial landmarks, which draws inspiration from wide applications of License
Plate Recognition (LPR). These landmarks are embedded with characters and
digit numbers that are similar to the name plates in offices and the license plates
used in transport. A similar approach is presented in Reference 21, which pro-
posed a visual landmark learning and recognition system for use in mobile
robot navigation tasks that can read text inside well-defined landmarks such as
nameplates, streets, and roads. However, there is no indication of its real-time
performance.
Figure 4.5a presents the format of the proposed landmark, and Figure 4.5b
shows a real landmark held by a person. Each landmark has the following
features:

• Five characters, the letter L followed by four digits, are printed on
the landmark.
• Each of the five characters has the same size, and the clearances
between the characters are all the same (H, W, and D in Figure 4.1a).
We currently select the parameters: L = 33, D = 200, H = 66,
and W = 34 (mm), which may be changed in different application
environments.
• The positions of the characters are also known (L in Figure 4.5a).

4.4.1 Landmark Recognition
The digits are the index of the landmark and the algorithm can identify the
landmark with a digits recognition method. The standard size of the charac-
ters contains enough information for robot localization. Since the proposed
landmark is similar to a license plate, many algorithms developed for license
recognition can be used here directly, including the fuzzy-map method for
locating the plate and the neural network for character recognition [22], and
the fast plate location method based on vertical edges of the images [23].
Figure 4.6 shows a new landmark recognition algorithm that consists of three
major modules: region finding, digits finding, and digits recognition.

4.4.1.1 Region finding module
This module is to find out all the probable regions that contain the landmark
digits and exclude as much background as possible. Considering the features
of the digits (sharply rising and falling edge in pairs in a horizontal scan line),

© 2006 by Taylor & Francis Group, LLC

FRANKL: “dk6033_c004” — 2006/3/31 — 16:42 — page 161 — #13