Computational Intelligence - November 2013 - 60

PanCam
Stereo WideAngle Cameras
(WAC)
Position
HighResolution
Camera (HRC)
Position
Aerial or
Satellite
Position

GIDS

VDF

3D Point
Extraction

LOD

Object
Detection

3D Data
and Image
Fusion

SLAM

GIS
Includes
3D and
Grayscale

Object
Recognition

SLA/EML

3D Point
Extraction

Updated
Navigation
Parameters

Path
Planner

Figure 14 Shows the autonomous and computational intelligent MSDP architecture.

PanCam
and
Multiple
Sensor
Inputs

3D
Points
2D
Images
Rover
Sensor
Data
MissionSpecific
Sensor
Data

Shape,
Roughness,
Obstacles

MSDP
GIDS

Texture,
Color

Appearance

Temperature,
Pressure

Environment
and Rover
Interaction
MissionSpecific
Environmental
Features

EML

Water
Vapor,
Radiation

Autonomous Data Processing via Learning

Path
Cost-Map
EML
via
Learn

RealTime
SLA

Figure 15 Shows the rover autonomous data processing via learning.

of the object in question, so as to increase
the identification speed of similar objects
and to reduce the load on computational
resources. The main challenges here are to
achieve learning in real time and to
improve object detection speed and
robustness. The existing data mining and
learning techniques applied in other applications such as search engines are relevant
but not readily or directly applicable.
Hence, for the benefits of the design of
SLAM and LOD, a self-organizing map
with neural networks adequately classifies
the sensor outputs during the learning process [12]. Using geometric hashing for the
abstract features in 3D space [18], the neurons in the lattice generate the self-organizing feature map. According to the topolog-

ical order, the learning algorithm of the
self-organized map is outlined as follows:
1) Randomly select a weighted value
from the neurons.
2) Pick three sample points from the
3D space.
3) Matching through all neurons
according to the given weighted
value to find the winning neurons.
4) Update the weighted value from
the winning neurons.
5) Iterate the process from step (2)
until it becomes a steady state.
Most importantly the learning rate of
SLA depends on the size of kernel that
we convolve to find the features of curvature class and torsion class within the
neighborhood of local surface. Similarly,

IEEE ComputatIonal IntEllIgEnCE magazInE | novEmbEr 2013

this applies to the other classes such as
roughness, obstacles, texture, color, temperature, pressure, water vapor and radiation as defined in Table 6 and Table 7.
D. Environment Model Library (EML)

EML consists of detected objects and environment properties that are mapped with
the location information from the GIDS.
The library is the result of multiple pipeline processes [8][23] shown in Figure 15
and provides a solution for autonomous
navigation [4][11] and environment perception. Objects can be abstracted in various classes based on; for example, geometry, appearance and material attributes. As
the autonomous rover interacts with the
environment, the library offers a database
of human-like interpretation of the environment. This in turn allows the rover to
make human-like decisions [26], such as
producing a path cost-map that is optimized based on the EML and can be used
for path planning of the rover as given in
Figure 16. SLA is responsible for comparing the partial information obtained during the detection stage with data within
the EML to achieve classification. If
required, data from the EML can also be
displayed from a 3D virtual reality simulator at the ROCC such as the one presented in [27]. This concept can increase
learning accuracy by incorporating human
feedback in a timely manner.
VI. Conclusions and Future Research

Our investigations in the Himalayas and at
Mount Everest provided useful information for PanCam and for future R&D.
Capturing images according to the ExoMars rover baseline information in the
RSM gave us the opportunities to
explore how the planning works when
the rover interacts under hostile environment. Mount Everest provides a good
analogue for an excursion on the surface
of Mars [10]. For example, using the captured WAC stereo images we have successfully re-constructed the 3D stereo
images of Mars-like-landscapes at EBC.
Furthermore, we demonstrated that the
data volume can be significantly reduced
with minimum loss in its image quality
after sub-framing, data compression and
super-resolution. Finally, the novel MSDP

Table of Contents for the Digital Edition of Computational Intelligence - November 2013