Page 52 - Autonomous Mobile Robots

Visual Guidance for Autonomous Vehicles 35

sequences. It is attractive because it avoids the requirement for a priori models of
the environment. The techniques are based on the constraints that exist between
the multiple views of features. This is a mature area of computer vision that
has attracted intensive research activity in the previous decade, prompted by
the breakthroughs in multiple view geometry in the early 1990s. Much of the
original work was motivated by mobile robotics but soon found more general
application such as: the generation of special effects for cinema, scene recovery
for virtual reality, and 3D reconstruction for architecture. Here, the theoretical
drive has been inspired by the recovery of information from recorded sequences
such as camcorders where the motion is general and little can be assumed
regarding the camera parameters. These tasks can be accomplished off-line and
the features and camera parameters from long sequences solved as a large-scale
optimization in batch mode. As such, many would regard this type of SFM as a
solved problem but the conditions in vehicle navigation are specific and require
separate consideration:

• The motion is not “general,” it may be confined to a plane, or
restricted to rotations around axes normal to the plane.
• Navigationisrequiredinreal-timeandparametersrequirecontinuous
updating from video streams as opposed to the batch operations of
most SFM algorithms.
• Sensory data, from sources other than the camera(s), are usually
available.
• Many of the camera parameters are known (approximately)
beforehand.
• There are often multiple moving objects in a scene.

Visual guidance demands a real-time recursive SFM algorithm. Chiuso
et al. [42] have impressive demonstrations of a recursive filter SFM system that
works at a video frame rate of 30 Hz. However, once we start using Kalman
filters to update estimates of vehicle (camera) state and feature location, some
would argue that we enter the already very active realm of simultaneous local-
ization and mapping (SLAM). The truth is that there are differences between
SLAM and SFM and both have roles in visual guidance. Davison [43] has
been very successful in using vision in a SLAM framework and Bosse [9] has
published some promising work in indoor and outdoor navigation. The key
to both of these is that they tackle a fundamental problem of using vision in
SLAM: the relatively narrow FOV and recognizing features when revisiting a
location. Davison used active vision in Reference 4 and wide-angle lenses in
Reference 43 to fixate on a sparse set of dominant features whereas Bosse used
a catadioptric camera and exploited vanishing points. SLAM often works well
with 2D ladar by collecting and maintaining estimates of a sparse set of features
with reference to world coordinate system. A problem with SFM occurs when

© 2006 by Taylor & Francis Group, LLC

FRANKL: “dk6033_c001” — 2006/3/31 — 16:42 — page 35 — #35