Advanced Image Understanding and Autonomous Systems

The ultimate goal of most image understanding systems is to produce an unambigu ous D representation of the local visual environment This representation can then be employed by robotic systems to e ect some meaningful action A great deal of research e ort is concerned with the development of visual and manipulative representations and their generative processes which allow for the e ective linking of such visual perception and robotic action Part of the motivation for this e ort is the desire to develop au tonomous systems It is argued in this paper that the requirements for the development of autonomous systems are not fully compatible with the current representation based A I paradigm While this approach is ideal for the construction of goal oriented systems which function in environments that can be speci ed a priori it does not and cannot address the problems encountered when adaptive self determining autonomous systems are required It is argued that such autonomous systems must be self organizing The problem which then arises is how such autonomous systems can be imbued with a goal oriented behaviour which re ects the requirements of its designer This remains an open question Current Approaches to Image Understanding Image understanding systems in general and robot vision systems in particular are concerned with the automatic interaction between computer based machines and their environment This interaction is facilitated by on going intelligent interplay between perception on the one hand and action be it navigation or manipulation on the other This is perhaps best characterized by the currently popular paradigm of active vision where the sensor is actively moved through the local environment to validate and re ne the interpretation the system has formed of the scene before it Current approaches to image understanding are based for the most part on the assumption that if the image understanding system is concerned with its environment it must somehow abstract relevant information about the environment so that it can reason with it These two aspects of vision the representation of information and the processes which facilitate the abstraction of that information form the kernel of current This work was supported by the Hitachi Dublin Laboratory their assistance is gratefully acknowledged vision systems Most advanced image understanding systems utilise several mutually relevant information representations based e g on the object edges or boundaries the disparity between objects in two stereo images and the shading of the objects surface and incorporate di erent levels of representation in order to organise the information being made explicit in the representation in an increasingly powerful and meaningful manner Typically an image understanding system will endeavour to model the scene with some form of parameterised three dimensional object model built from several low level processes based on distinct visual cues The tenets of conventional computer vision systems fall soundly in the domain of representationalism i e the philosophy that perception is a mechanism by which the entity apprehends the world in which it nds itself learns its structure and models it and modi es its behaviour on the basis of what it learns If we accept for the moment the va lidity of this approach then a number of questions arise What are these representations What processes are necessary to generate them How are these processes organised The answers to these questions form the body of research in advanced image understanding There is not su cient space in this paper to include an exhaustive review of each of these processes and representations Indeed to do so would distract us from the central argument in the paper which concerns the relationshop between image understanding systems and autonomous systems A summary can be found in and in and we will instead provide a brief thumbnail sketch of a few representative visual representations and processes Organisation of visual processes In proceeding from raw D images of the world to explicit D structural representa tions we are making a signi cant leap across widely divided levels of representation the information inherent in the former is implicit and iconic that in the latter is explicit and predominantly symbolic To traverse this gap we must accept that no single pro cess nor representation is going to be generally adequate Consequently a central theme which runs through the current conventional approach to image understanding is that intermediate representations are required to bridge this gap between raw images and the abstracted structural model This realisation owes much to the work of David Marr see who exerted a major in uence on the development of the computational approach to vision Marr modelled the vision process as an information processing task in which the visual information undergoes di erent hierarchical transformations at and between lev els generating representations which successively make more and more three dimensional features explicit These representations make di erent kinds of knowledge explicit and should expose various kinds of constraint upon subsequent interpretations of the scene It is the pro gressive integration of these representations and their mutual constraint to facilitate an unambiguous interpretation of the scene that most characterises this approach to vision We can characterise image understanding then as a sequence of processes concerned with successively extracting visual information from one representation beginning with digital images organising it and making it explicit in the representation to be used by other processes From this perspective vision is computationally modular and sequential Visual Representations