Sun Annual Lecture 1995
***********************

Teaching Vehicles to See
========================
by Ernst D. Dickmanns 
======================

31st August and 1st September 1995
++++++++++++++++++++++++++++++++++

 o About the Sun Annual Lecture 
 o Synopsis 
 o About the lecturer 

About the Sun Annual Lecture
============================

The Sun Annual Lecture in Computer Science at the University of
Manchester provides an opportunity for eminent computer scientists to
give a series of lectures introducing a wider audience to current
research work in their area. The lectures occupy about eight hours
over two days, giving ample time for questions and discussion.

Registration for the lecture series is 90 pounds (45 pounds for
registered students), which includes lunch each day, and supporting
material for the lectures. To register, print the registration form,
fill it in, and return it to the address shown, together with your
payment. Registration is free for members of the University of
Manchester, who can register by e-mail or telephone (see below) or by
printing and returning the alternative registration form.

The Annual Lecture dinner will be held on the night of the 31st at
Yang Sing, one of the best Chinese restaurants in the country. If you
would like to attend this, please tick the box in the registration
form and include the cost with the registration payment. Please
indicate if you are vegetarian or have other dietary preferences.

We can arrange bed and breakfast accommodation at St. Anselm Hall (a
student hall). Please indicate on the registration form if you want us
to make bookings for you. Note that payment for accommodation is made
separately to the Hall before departure.

For further information, please contact: 

   Jenny Fleet,
   Annual Lecture,
   Department of Computer Science,
   The University,
   Oxford Road,
   Manchester M13 9PL,
   ENGLAND.
   0161-275 6130 
   annual-lecture@cs.man.ac.uk
   http://www.cs.man.ac.uk/events/sun-lecture.html


Synopsis
========

Each lecture will last approximately one hour. 

Introduction 

Emphasis is put on the task context for machine vision since this
determines the side constraints dominating visual interpretation and
hypothesis generation for ecological dynamic scene understanding. In
addition to 3-D space, the dimension of time is thoroughly considered
for making image sequence analysis efficient by using known invariants
of objects in all four basic dimensions and for preventing
combinatorial explosion of feature grouping. Time derivatives and
their integrals are exploited in a systematic way, both for perception
and for control; for the general case of observing moving objects with
cameras on board a moving vehicle, inertial sensors like
accelerometers and rate sensors are very important for high
performance dynamic vision. In this context, perception means, more
precisely, the combined use of inertial, odometric and image sensors. 

   This lecture will include videos of road and air vehicle guidance 


Task analysis
   Typical sets of tasks. The basic four dimensions (4-D): 3-D space
   and time.
   Coordinate systems, resolution scales, multiple scale modelling for
   different
   aspects of the vision task. Environmental conditions and objects:
   representation according to the visual task. The body and its
   control: mission
   decomposition (mission elements, types and sequencing); perceptual and
   behavioural capabilities required; functional realizations through
   feedback and
   feedforward control; symbolic representations of behavioural capabilities.
   Examples: road vehicle guidance, aircraft landing approach. 

The 4-D approach to dynamic machine perception and control 
   Generic object and process models for feature based machine vision.
   Differential and integral representations in space and time. The
   idea of Gestalt
   (shape and aspect conditions). The instantiation problem and the initial
   orientation phase. Fast recursive updates through prediction error feedback
   (Kalman filters and derivatives). Model based integration of image and
   inertial measurement data. Modular overall system design. The dynamic data
   base (DDB) for object-oriented exchange of actual best estimates. Active
   viewing direction control and attention focussing. A survey of control
   computation. 

Sensory inputs and data interpretation 
   Active bifocal vision for both a large field of view and good
   resolution in the
   region of special interest. Parameter selection. Data versus information.
   Features as carriers of information. The KRONOS software package for
   edge-feature extraction. Efficiently controlled feature extraction through
   temporal predictions. Bottom-up versus top-down interpretation schemes.
   Object processor groups for confining communication requirements. Inertial
   and other conventional measurements. Object- and expectation-based data
   fusion. Perceptual organizations and resulting capabilities. Inertial
   stabilization, saccades and smooth pursuit as active vision
   modes. Application
   examples in road vehicle guidance. 

Vision system integration and task performance 
   Vehicle control by a 3-layer hierarchy for decoupling fast reflex-like
   reactions from event-triggered manoeuvre elements and knowledge based
   behavioural decisions. Examples from road vehicle guidance: lane
   and distance
   keeping by feedback control; lane changes by feedforward control time
   histories; superposition of visual feedback components for handling
   perturbations; decisions for lane change by situation analysis and
   assessment in
   the task context; mode transitions and monitoring of actual behaviours. 

Example: road vehicle guidance 
   Coherent discussion of visual road vehicle guidance tasks. The test vhicles
   VaMoRs (a five ton van) and VaMoRs-P (VaMP for short, a passenger car),
   their equipment and performance levels. Motorway driving at high speeds
   with multiple objects to react to; the approach to multiple object
   detection and
   tracking. Driving on state roads and negotiating turn-offs onto cross roads
   including active vision strategies; region-based feature extraction with the
   `triangle' algorithm. Driving on unsurfaced minor roads and
   recognizing forks
   in the road. Visual recognition of vertical curvatures with low and
   high spatial
   frequency components. 

   There will be video of these vehicles in action. 

Full 3-D motion: landing approaches by dynamic machine vision 
   Modelling of perspective projection with constrained motion in all
   six degrees
   of freedom (3 translations and 3 rotations) for visual recognition
   of the state
   relative to the landing strip. Fixing the view relative to a point
   on the horizon
   for easier data interpretation. Initialization through systematic search.
   Selection of image evaluation areas (windows) for high frequency recursive
   data interpretation. Handling perturbations from turbulence using inertial
   measurement data. Delay-time compensation for visual data interpretation.
   Transputer hardware for dynamic machine vision. Hardware-in-the-loop
   simulation results, and flight experiments with a twin turbo-prop Do-128. 

Visual grasping of a free-floating object in space, and conclusions 
   In May 1993, the Space Shuttle/Spacelab-D2 Robot Technology Experiment
   (ROTEX) was set up in low Earth orbit with a camera in the robot arm and the
   computing facilities on the ground in Oberpfaffenhofen near Munich. This
   was the first machine tele-operation experiment, with a large time delay of
   about six seconds. The lecture will discuss man/machine interaction for
   initialization, automatic delay compensation and tele-grasping. 

   Lines of development for dynamic machine vision. In the near future,
   computing power will be sufficient for solving practical tasks with
   affordable
   vision systems, but robustness has yet to be improved. 



About the lecturer
==================

Professor Dr.-Ing. Ernst D. Dickmanns is currently Professor for
Control Engineering in the Department of Aerospace Engineering at the
Universitaet der Bundeswehr in Munich. He began his studies in
aerospace engineering at RTWH Aachen from 1956 to 1961. He then joined
the German National Aerospace Research Establishment (DFVLR) in
Muelheim-Ruhr and Oberpfaffenhofen as a researcher in the field of
optimal control and trajectory shaping. From 1964 to 1965 he undertook
graduate study at Princeton University on a NASA fellowship, and on
his return to DFVLR became head of the trajectory dynamics section of
the Institut fuer Dynamik der Flugsysteme. He received his doctorate
from RTWH Aachen in 1969.

>From 1971 to 1972, he held a postdoctoral research associateship with
NASA where he investigated Space Shuttle Orbiter design. From 1972 to
1974, he was in charge of designing the transfer and positioning
procedures for the first European geostationary satellite,
Symphonie. In 1974, he became acting head of the DFLVR Research Centre
in Oberpfaffenhofen.

He took up his current appointment at the Universitaet der
Bundeswehr in 1975, where he founded the Institute for System Dynamics
and began the research programme on real-time machine vision for
autonomous vehicle navigation. This has undertaken pioneering work in
visual dynamic scene understanding and real-time autonomous visual
guidance for road, air and space vehicles. Under his direction, seven
autonomous road vehicles have been equipped with vision systems, of
which four are currently in operation in normal traffic on the German
autobahns. They have accumulated over 6000 km of fully autonomous
driving with both longitudinal and lateral degrees of freedom. Similar
techniques have been applied to aircraft landing approaches, grasping
of a free-floating object in satellite orbit, and helicopter control.

Last updated 9/5/95 by Tim Clement (timc@cs.man.ac.uk)