Active vision and adaptive learning

Active vision is identified by a closed loop linking sensing with acting. Thus, an active vision system's behaviour is directly determined by what it senses. To date however, the responses produced by active vision systems have tended to be relatively low-level, generally designed to facilitate improved sensing, by enhancing the duration or speed of object tracking, for example, or optimising the focussed application of more intensive image processing. This is probably adequate if the active vision system is designed as a front end to other processes or to specialised application systems, or if it is a demonstration in support of a theoretical vision model.
    However, this leaves unanswered the problems of i) how to select an appropriate action when many different alternatives are available, and ii) how best to modify the behavioural repertoire of the system. These problems are especially important in two situations: firstly, when an autonomous system faces a novel situation and must respond adaptively without the benefit of a priori knowledge, and secondly, when systems attempt higher levels of perception and response, and links between the absolute properties of the incoming image data and the actual objects of perception become increasingly attenuated.
    This paper discusses methods for linking learning with active vision so that the behaviour of the system is optimised over time for the achievement of goals. We argue the necessity of system goals in learning vision systems, and discuss methods for propagating goals through all levels of loose hierarchies. In the last section we outline an architecture in which high and low level perception operate interactively and in parallel.

Mark W Peters and Arcot Sowmya, 1996.

Intelligent Robots and Computer Vision, SPIE, Boston.

WRAITH: Ringing the changes in a changing world

This paper documents techniques we have used to control the attention of a visual robot assigned the task of following (in the sense of both tracking and understanding) human movement. We believe that even simple tasks like intelligently controlling the direction of gaze must involve a complex synthesis of all levels of perception, from those of ‘early’ vision to sophisticated socio-cultural levels. In order to produce real-time real world intelligent behaviour, two key questions must be answered: how are such levels of perception generated in the first place, and how do they combine to create coherent, useful behaviour. We explain how we have approached these problems and present some results of current work in the WRAITH project. The paper also discusses our hierarchical conceptual framework of motion perception, which has guided the work. We suggest that, in addition to having the essential faculty of pattern recognition, intelligent systems must detect and react to pattern change, often before they have had time to identify the patterns. In other words, it is just as important to ring the changes in the world as it is to notice its regularity.

Mark W Peters and Arcot Sowmya, 1997.

4th Conference of the Australasian Cognitive Science Society, Newcastle, Australia.

A real-time variable sampling technique for active vision applications: DIEM
ICPR 1998 Best student paper prize winner

We describe a sampling technique particularly suitable for active vision: Dimensionally-Independent Exponential Mapping (DIEM), in which each dimension of the original data is sampled in an exponentially increasing or decreasing series of steps, with bilateral symmetry about the data mid-point. Multidimensional data sampling is achieved by combining single dimension sampling co-ordinates. DIEM is simple, fast, flexible and very useful for active vision, but may also have applications in other domains, possibly of higher dimensionality. Its most unusual feature, invertibility, is also one of its most useful. The many advantages of DIEM (computational simplicity, preserved linearity, real-time adjustability, learnability, human motion suitability, expandability, invertibility, frame grabber compatibility, and affordability) are each described. We also describe the functional characteristics of DIEM, provide formulae for DIEM specification and verification, and refer to how DIEM may best be exploited, giving our own work in visual robotics as an example.

Mark W Peters and Arcot Sowmya, 1998.

International Conference on Pattern Recognition '98, Brisbane, Australia.

Spatial competence via self-organisation: an intersect of perception and development

We address the question of how artificial systems and natural organisms develop spatial competence. Most artificial systems draw upon considerable sophisticated operator- or developer-originated knowledge about what in the world sensor signals represent. Natural systems do not have such sophisticated auxiliary sources of information. We are interested in how, despite this, they achieve perceptual organisation, and suspect that the methods they use will have generalisable effectiveness. We describe a process that creates coherent mappings between the physical world and the phenomenological realm, analogous to retinotopicity and sensory homuncularity in natural systems, and discuss its application to problems of higher dimensionality and higher levels of abstraction. Importantly, such a process, having proved successful in the perceptual robotics domain of our current interests, is likely to be found in other cognitive domains because its strengths lie in its ability to organise and implicitly summarise data in the absence of clues about what that data represents.

Mark W Peters, 1998.

Cognitive Science '98, Madison, Wisconsin.

Towards artificial forms of intelligence, creativity, and surprise

This paper starts out from two observations: firstly, that there are complex links between what we term intelligence and what we term creativity and, secondly, that the phenomenon of surprise has a significant role in both the genesis and evaluation of creativity, and is tightly coupled to perception. We argue that for machines to develop to the point where we attribute to them intelligence and, therefore, their own degree of creativity, they must first develop a sensibility of surprise. This, we show, is predicated upon a multi-level organisation of perception, and a method of representing the interest, or novelty, of events and actions taking place in the physical world. A sensibility of surprise further depends on an ability to recognise the novelty of actions the system itself is contemplating. We describe methods of encoding surprise in perceptual robots, and show how this enables them to focus on what is interesting in their environment – a prerequisite to the production of behaviour both creative and intelligent.

Mark W Peters, 1998.

Cognitive Science '98, Madison , Wisconsin.

Integrated techniques for self-organisation, sampling, habituation, and motion-tracking in visual robotics applications

We summarise several techniques in use in our visual robotics research. Our aim is to develop robots that are thoroughly autonomous and adaptable. We describe a system that is independent of typical image derivation standards, sampling algorithms, a priori space, object, or motion models, yet is able to intelligently direct its attention to novel activity in a complex and changing environment.

Mark W Peters and Arcot Sowmya, 1998.

Machine Vision Applications '98, Chiba, Japan.
 

Making sense: autonomy and adaptation in visual robotics

Preliminaries

Chapter 1, Introduction

Chapter 2, Background

Chapter 3, Constructing a Framework - WRAITH

Chapter 4, Organising the Sensorium - JIGSAW

Chapter 5, Selecting Data - DIEM

Chapter 6, Developing Sensibility - SURPRISE

Chapter 7, Discussion

Appendices

Addenda  

This is a practical and theoretical thesis in visual robotics. It describes the development and actual implementation (in a real world, real-time visual robot) of new approaches to three non-linear problems: how to ensure robustness under the harshest conditions of unforeseen reconfiguration, how to provide specialised space-variant sampling regimes according to which task is currently at hand, and how to automatically direct attention using any number of adaptive response layers in concert. Additionally, the descriptions of this practical work are preceded by the exposition of a new theoretical framework for intelligent system evaluation, which offers performance silhouettes as a schematic method.
    The three practical methods stem from, and are embedded in, the theoretical framework, and all have been suggested, to varying degrees, by knowledge of biological processes or capabilities - self-organisation, visual periphery sensitivity, and adaptive reduction in sensitivity, for example.
    The overall goal of the research is to develop extremely simple algorithms capable of operating in real-time and endowing a robot with robust essential perceptual capabilities that can operate in all environments.
    The outcome of the work is a visual robotic system that exhibits seemingly intelligent behaviour in complex, changing, and noisy natural environments. That it does so with minimal help from other sophisticated agents (such as computer science researchers) is a credit to its autonomy and the adaptation of its design to arbitrary environments.
    The unifying theme is the development, ultimately, of a set of algorithms that can be dropped into an unknown sensor-motor apparatus, then wake it up and embark on a series of input-output experiments, resulting in the apparatus's ability to understand itself, and thus understand the world; control itself, and thus control the world.
    The major conclusion to be drawn from this work is that radically autonomous systems can be built, and will work in the face of traumatic physical insults that would destroy the performance of ordinary machines. In this respect, they can exhibit the adaptation and graceful degradation of performance so often admired in biological systems.

Mark W Peters, 2000.

PhD Thesis.

Jigsaw: the unsupervised construction of spatial representations

A fundamental assumption in machine vision is that the spatial arrangement of pixels is given. In challenging this assumption we have utilised a general relationship that exists between space and behaviour. This relationship presents itself as spatial redundancy, which other researchers have considered problematic. We present a mathematical model and empirical investigations into this relationship and develop an algorithm, JIGSAW, which uses it to build spatial representations. The philosophy underpinning JIGSAW takes signal behaviour, rather than position, as primary. JIGSAW is an unsupervised learning algorithm that is efficient in time and space and that makes minimal assumptions about its operating domain. This algorithm offers engineering potential, opportunities in the understanding of biological vision, and a contribution to the wider field of cognitive science.

Mark W Peters and Barry Drake, 2000.

Technical Report, UNSW-CSE-TR-0007, University of New South Wales.

Fundamental robust self-organising vision

This paper presents JIGSAW, an algorithm that performs a fundamental constructive sensory process. It is self-organising robust, compact, and similar in function to the basic organising processes that take place in pre-natal neuro-visual development. JIGSAW solves the problem of how to autonomously build a spatial representation without spatial precedents or models. The key function of JIGSAW is to turn a population of unordered independent inputs, each bearing a signal from a sensor of unknown position, into a unified spatial representation that maps the actual environment being sensed. This function is important because it demonstrates the emergence of a completely new perceptual level, and because it provides a fully adaptive platform for further visual processes. As spatial competence is an important attribute for animals and robots alike, this is a useful method. JIGSAW is robust because it does not depend on its data being spatially organised. It can be interrupted and restarted at any point, without prejudice to the eventual result. Results from five experiments, each with a different kind of data, are presented and discussed.

Mark W Peters and Barry Drake, 2000.

4th Australia-Japan Joint Conference on Intelligent and Evolutionary Systems, Hayama, Japan.

Mathematics and Celtic design (2nd edition) 

Celtic design interests me. I find the methods employed in the construction of designs fascinating for their own values, but also useful as a foundation for some of my own work and a source of parallel mathematical principles. I shall not endeavour to prove that there is mathematics in Celtic design, but rather that mathematics offers a present-day homology of it.
    I have always sought to understand as simply as possible that which interests me and, from my own point of view, understanding is most expedient if it is in a practical form, that is, giving the capacity to utilise knowledge to create designs that, if necessary, will fulfil the requirements of Celtic work.

Mark W Peters, 1979/2001.

BA Honours Thesis.

other pages by
other pages
back to main page
home page
information page