Raj Madhavan, Ph.D.


The Intelligent System's Division of the National Institute of Standards
and Technology (NIS) has been developing control systems architectures,
advanced sensor systems, research services, and standards to achieve
autonomous mobility for unmanned ground vehicles. Presently, the
eXperimental Unmanned Vehicle (XUV) can autonomously navigate at 60 km/h
on-road and at 35 km/h off-road in daylight, and 15 km/h off-road at
night or under inclement weather conditions by employing the
4D-Real-time Control System (4D/RCS) hierarchical architecture for
autonomous navigation. The primary navigation sensor suite of this
vehicle consists of a Schwartz Electro-Optics LADAR, color cameras,
Global Positioning System (GPS), and an Inertial Navigation System
(INS). The navigation system uses a Kalman filter to fuse data from the
INS and the carrier phase differential GPS unit to compute the vehicle's
position, orientation, speed, and velocity. GPS (even in differential
mode) is not always reliable as the accuracy of the GPS estimates
depends on the direct line of sight between the GPS receiver and the
number of satellites. Hence to continually provide reliable vehicle
information, an alternate solution becomes inevitable.

With increasing computing power and reduction in the size and increase
in sophistication of LADAR units, our motivation to register 3D LADAR
data (both consecutive range images and also range images to a priori
maps) is many fold: In addition to eliminating the dependence on GPS,
LADAR registration can serve either as a primary navigation solution or
as a secondary combined solution along with GPS. Within RCS, the use of
a priori maps would enhance the scope of the world model. These maps may
take a variety of forms including survey and aerial maps and may provide
significant information about existing topology and structures. In order
to take advantage of this knowledge, research is needed to register
these a priori maps with the sensor-centric maps. Additionally, for
incorporating a priori knowledge into the world model, some form of
weighting is required and this depends on how well the a priori data and
the sensed information are registered. There is also the need to
generate higher resolution a priori terrain maps, as the current survey
maps are too coarse for autonomous driving. Another potential
application for registering LADAR data is the computation of ground
truth as such registration is not dependent on time-based drift (unlike
INS), vehicle maneuvers and terrain of travel. All of the above
mentioned factors warrant the need to develop robust 3D small LADAR data
registration algorithms.

To register 3D LADAR data obtained vehicles moving in an unstructured
and unknown environment, an iterative registration algorithm has been
developed. The proposed is iconic as it works directly on sensed data
and thus does not require explicit extraction and matching of features
and experimental results have shown that it is robust to gross outliers
and occlusions in the data. The enclosed figures fig01 and fig02
show two sets of range images before and after registration, respectively.