Research Projects
 Model Selection
Large-Scale 3D Vision
Sensor Data Fusion in Brake-by-Wire Technology
Robotic Aircraft
Optical Flow Calculation
Motion Segmentation
Range Segmentation
Model Selection
Problems in Model
Selection for Range Segmentation
The available
information theoretic model selection criteria are based
on the assumptions that noise
is very small and the data size is large
enough. However, in visual data segmentation
algorithms, one usually deals with noisy
regions of data, some of which are small
localised regions that are to be extended
later. As a result, neither of the above
assumptions is usually valid.
The other crucial problem is that almost all
the information-theoretic model selection
criteria measure the complexity of a model
only by its number of parameters1. Hence, if there are different models
with the same number of parameters (i.e.
same complexity), there will be some difficulty
in selecting between those models. In fact,
the available model selection information-theoretic
model selection criteria assume the set of
candidate models to be nested, whereas, in
many applications, such as range segmentation
of curved objects, the set of candidate models
may consist of non-nested models.
Finally, all the
available model selection methods have their
roots in statistics, the physical and geometrical
characteristics of the problem have been
ignored.
As a result none of
the existing model selection criteria is capable
of identifying the true underlying model for visual
data in motion segmentation and range segmentation
applications.
Surface Selection Criterion
To overcome the above
problems in model selection, we proposes to view
the sum of bending and twisting energies of a surface
as a measure of the surfaces roughness and the
sum of squared residuals as a measure of fidelity
to the true data. There should be an acceptable
compromise between these two factors. To formulate
the proposed Surface model Selection Criterion
(SSC), different points of a surface are viewed
as hypothetical springs constraining the surface
as shown in the below figure.
If the surface has
little stiffness, then the surface bends and twists
to meet the points. Hence, the surface fits itself
to noise and the sum of squared residuals between
the range measurements and their associated points
on the surface will be small. This bending and
twisting in turn increases the amount of strain
energy accumulated by the surface.
|
Representation a malleable surface
supported by hypothetical springs. The range measurements
are shown by black circles. |
As one may expect,
increasing the number of parameters of a surface
leads to larger bending and twisting energies
as the surface has more degrees of freedom and
consequently the surface can be fitted to the
data by bending and twisting itself so that a
closer fit to the measured data results. This
can be inferred from the bending energy formula
Equation 1). However, the higher the number
of parameters for a surface model, the less the
sum of squared residuals. For instance, in the
extreme case, if the number of parameters is
equal to the number of data points (which are
used in the fitting process), then the sum of
squared residuals will be zero whereas its sum
of energies will be maximised.
If a plate is bent
by a uniformly distributed bending moment so
that the xy and yz planes are the principal planes
of the deflected surface, then the strain energy
(for bending and twisting) of the plate can be
expressed as:
 (1)
where D is the
flexural rigidity of the surface and ν is Poisson’s ratio (ν should
be very small because in real world-objects the
twisting energy in comparison with the bending
energy is small). In all the experiments reported
in this thesis, n is
assumed to be 0.01. As our experiments show,
the performance of SSC is not sensitive to the
small variation of n .
The strain energy is computed as a measure of
complexity of the model.
In order to scale
the strain energy, it has been divided by the
strain energy of the model
with the highest number of parameters (Emax).
Therefore, D has been eliminated from the following
formulations.
To establish the
trade-off between the sum of squared residuals  and the strain energy EBending+Twist,
a function SSC is defined such that:
where δ is the scale of noise for the highest surface (the surface
with the highest number of parameters). The
reason for using the scale of noise of the
highest surface (as explained by Kanatani [1])
is that the scale of noise for the correct
model and the scale of noise of the higher
order models (higher than the correct model)
must be close for the fitting to be meaningful.
Therefore, it is the best estimation of the
true scale of noise, which is available at
this stage.
An accurate estimate
of the scale of noise δcan
be computed by  where N is the number of data points and P is
the number of parameters of the highest surface.
Use of this formula for the scale of noise can
be justified by the fact that if the model is
correct, then  is
subjected to a χ<2 distribution
with N-P degrees of freedom. The energy
term has been multiplied by the number of parameters P in
order to penalise the choice of a higher order
(than necessary) model. Such a simple measure
produces good discrimination and improves the
accuracy of the model selection criterion.
Having devised a
reasonable compromise between fidelity to data
and the complexity of the model, the model selection
task is then reduced to choosing the surface
that has the minimum value of SSC.
Large-Scale 3D Vision
This research project is intended
to provide the necessary resources for comprehensive large‑scale
three‑dimensional mapping of man‑made environments.
Achieving a complete 3‑D digital representation of the surrounding
environment has, over the years, received substantial attention
due to the plentifulness of its applications.
To achieve the goal of large-scale
3-D environmental mapping, a 3-D range scanner has been developed.
The equipment is capable of producing range maps of far away objects
(up to 300 meters) with relatively wide field-of-view. More importantly,
range and intensity images produced by this equipment can be fused
to have a complete set of (x, y, z, R, G, B) sextuplets of
the scene.
The range scanner system is based
upon a distance meter developed by MDL and a programmable pan-tilt
unit manufactured by Directed Perception. The rangefinder is capable
of producing distance measurement of 7-300 meters with repetition
rate of 1000 Hz. The typical accuracy of the machine is 300 mm,
which is 0.1% of the maximum measured distance.
In this project we mainly focus on
development of an accurate and calibrated measurement device, which
its measurements are sufficiently accurate for a wide range of scientific
endeavours such as volumetric and surface recording of historical
sites, virtual reality in environment building and modelling and
architectural reproduction.
Following figures show an example
of our experimental result. Range image, taken by range scanner
device (left) and intensity image (middle), captured
by video camera, has fused (right) to present a rich data
set of the scene.
Sensor Data Fusion in Brake-by-Wire Technology
Brake-by-wire is a frontier technology
that will allow many braking functions to switch to electronic actuation.
It will lead to more effective and safe braking systems, elimination of
hydraulic technology, release of space and reduction of maintenance. By
the end of this decade, most international car manufacturers are envisaged
to have adopted brake-by-wire systems across their range of passenger
cars. A large variety of sensors are utilised in an Electro-Mechanical
Braking (EMB) system and therefore their consistent operation is vital
for the functionality of such system. To achieve a high level of coherency
amongst such a large collection of sensors (mandated by the safety requirement
of a brake system), the use of sophisticated data fusion techniques
is unavoidable. Data fusion theory is generally considered the most appropriate
tool to generate reliable and accurate information by virtue of combining
all the available data. During the last two decades, theory of data fusion
has been broadly developed and applied in many applications. Data fusion is
generally defined as “synergistic combination of the information, provided
by multiple sensory devices or multiple sources, to assist in the accomplishment
of a task by a multi-sensor system”. Generally speaking, data fusion techniques
are required in an EMB system for many important tasks that can be classified
into the following three categories:
-
Effective Utilisation of Redundant Information
An EMB, by nature, is a safety critical
system and therefore fault tolerance is a vitally important characteristic
of this system. As a result, EMB systems are designed in such way that
many of its essential information would be derived from a variety of sources
(sensors) and be handled by more than the bare necessity hardware. Three
main types of redundancy usually exist in an EMB system:
- Redundant sensors in safety-critical components such as the brake pedal
- Redundant copies of some signals that are of particular safety importance such as
displacement and force measurements of the brake pedal
- Redundant
hardware to perform important processing tasks such as central controllers
In order to utilise the existing redundancy,
voting algorithms and probability‑based fusion techniques need to
be evaluated, modified and adopted to meet the stringent requirements
of an EMB system. Reliability, fault tolerance and accuracy are the main
targeted outcomes of the fusion techniques that are developed especially
for redundancy resolution inside an EMB system.
-
Estimation and Identification
In a typical EMB system, there are three
categories of estimation and identification tasks that challenge the existing
sensor data fusion techniques. Those categories are:
- Accurate
and reliable estimation of environmental parameters (e.g. friction coefficient
between the tyre and the road) by integration of multiple sensors or sources
of data
-Parametric
and non-parametric identification of unknown dynamics in the system (e.g.
calliper dynamics)
-Estimation
of system states or variables by fusing multi-source information (e.g.
vehicle state estimation, estimation of the zero crossing point on the
actuator characteristic curve and clamp force estimation).
The central controllers
in the system are mainly responsible for high level braking functions
such as four-wheel proportioning, automatic hill hold, antilock brake
control, traction control, vehicle stability control and so on. Each
of these control tasks must be accomplished based on the data received
from multiple sensors and sources. Hence, robust and accurate multisensor
fusion methods are required to be developed if high level control tasks
of the EMB system are to be successfully executed.
We have been
working on different fusion problems that have arisen during the design
and testing of an EMB system. A new voting algorithm for fusion of the
brake pedal sensors, a novel predictive filter for compensation of the
missing samples of sensor data (filed as an Australian patent), and a
new angle tracking observer for estimation of the position and speed of
brushless DC motors based on a resolver sinusoidal signals (filed as another
Australian patent) have been developed sofar.
Robotic
Aircraft
 The main goal of
this project is to develop reliable computer vision
techniques to be utilised in the control of low altitude
manoeuvres for autonomous small-scale aircraft. In
particular, we intend to focus our attention on recovering
the relative altitude of an aircraft by using the
terrain texture seen from an on-board camera. We
aim to develop techniques that would work reliably
on almost any terrain and weather condition. We are
also committed to provide evaluation techniques capable
of verifying our estimates (as they are generated)
and associate a probabilistic reliability measure
to the sensor output.
Click here to
watch a clip which is taken by the aircraft.
 |
 |
The
model plane |
The
Engine Mounting |
 |
|
The
Camera Support |
|
Optical Flow Calculation
Differential optic flow techniques try to relate local
changes in image intensity (expressed as spatial
and temporal derivatives of the image brightness
function) to the optic flow. Differential techniques
usually perform faster than image matching or
phase based techniques and lead to a simple set
of linear equations. The fundamental assumption
behind all the optic flow techniques is the fact
that the brightness intensity function of a moving
object remains approximately constant at least
for short duration of time Optical Flow is the
motion of objects on the image plane as objects
(of the camera) moves. This has applications
in robotics and in video coding technology. Of
particular recent interest, is the study of motion
segmentation - how to break an image up into
the parts corresponding to objects undergoing
different motions. A significant factor in this
research, of late, is the use of Robust Statistics.
A major recent advance was the use of robust statistics to
calculate the flow in small patches (with which the flow
is assumed to be uniform) [Bab-Hadiashr & Suter
1998].
A (LARGE!) mpeg file
shows a screen capture of our optic flow demonstration
code -
 30Mb Mpeg file (created by Dr. David Suter)
Motion Segmentation
Segmentation is a vital task in many computer vision problems.
In our work we have concentrated upon applying our segmentation
methods to the segmentation of optical flow (motion) and range
data. However, our methods are sufficiently general and robust
as to apply to virtually any segmentation problem. Our focus
has shifted to motion segmentation to overcome the assumption
that the flow is uniform in certain patches and on the selection
of motion models. Major results in this direction can be found
in [Bab-Hadiashr & Suter
1999].
Following figures show an example of our approach to the
motion segmentation problem .
Otte sequence and its measured flow field
Segmentation Results - First and Last Stage
A (LARGE!) mpeg file shows a sequence segmentation
based upon motion.
17Mb Mpeg file (created by Dr.
David Suter)
Range Segmentation
Range data are
usually clean and accurate and although they are made of multi-structures
(at least in our examples), they do not have many impulsive outliers.
Thus, the range data segmentation task probably does not challenge
our methods as greatly as the motion segmentation task. However, clearly
there is still some utility in employing robust segmentation methods
on range data. Following figures show an example of our approach to
range data segmentation (for detail, see Bab-Hadiashr & Suter
1999).
  
Intensity image, original and segmented range
data
|
