Some of my work has a practical objective: the integration of control, reasoning, image understanding and real-time technologies for traffic surveillance systems. This has developed from my work on optic flow measurement techniques and motion interpretation algorithms, and now involves the higher level interpretation of motion patterns using stochastic and logic-based techniques. I am also developing perceptual control mechanisms for motion segmentation and model-based interpretation, and interested in parallel processing for vision and distributed control formalisms.
Keywords: Dynamic vision, visual surveillance, biomedical image analysis, human motion perception
My most important recent result has been a reformulation of the equations linking first-order optic flow (dilation, shear and rotation) to surface motion and orientation. This generalises the simple relationship between time-to-contact and dilation in the case of direct approach, and it clarifies our understanding of the way that first-order flow might be used for the control of locomotion both by animals and machines. I have demonstrated in simulation how the flow information plus knowledge of one additional variable is sufficient for control of a docking task in 3 dimensions. In addition, I have demonstrated the value of using log-polar sampling in first-order flow measurement, in the context of an active visual system. This has produced a fast and reliable method, which leads to the possibility of practical applications in real-time robotics.
Consideration of active vision and image representation inevitably leads to an interest in eye movements. Together with Hilary Tunley I have developed software for a video-based eye movement monitoring system using novel Hough transform and active contour techniques, and I hope in future to use this system to explore the use of motion information in humans.
Keywords: Optic flow, active vision, perceptuo-motor control, eye movements, image representation, mobile robots
For example, one aspect under development is an activity-based model of colour-processing in the early visual pathway. Principal component analysis of small (15x15) image fragments taken from natural colour images (see frog and parrot) shows that one or two of the main components are significantly colour selective but without orientation preference, whereas the rest are only selective to orientation. The result suggests an activity-based explanation of the formation of "colour blobs" in primary visual cortex. Subsequently we built a detailed network simulation of processing in the retina, lateral geniculate nucleus and adaptive visual cortex which self-organizes in response to natural colour images and produces "feature-maps" in which islands of a few unoriented colour-sensitive cells i.e. colour blobs are surrounded by a sea of oriented non-colour-selective cells (see receptive fields).
Neurons in the visual cortex, unlike those from the lateral geniculate nucleus (LGN) which convey retinal signals to the cortex, respond to, for example, bars and edges in the visual field of particular position, orientation and often motion. There is considerable debate about how these properties come about given our knowledge of the cortex.
The project attempts to understand how these properties might emerge through self-organisation (namely through Hebbian adaptation and the dynamic redistribution of connections) when an abstract 3D network of cortical cells is presented with either artificial or natural image fragments. Unlike previous models of this type we have tried to make the constraints on our network more realisitic in terms of the known biology. We employ, for example, substantial recurrent excitation (positive feedback), and relatively weak local inhibition.
Keywords: neocortex, early vision, adaptation, biological vision, neural networks
Keywords: Visual System Development, Hebbian Neural Networks, Learning Rules
Difference of Gaussian filtering and Gabor wavelet analysis is used to preprocess face images, mimicking the effects of receptive field functions found at various stages of the human vision system. These were then used as input representations to RBF networks that learnt to classify and generalise over different views.
These techniques have been extended from static frames to image sequences with collaborative work with QMW College, London, and have shown excellent generalisation in areas where data quality has been extremely poor. In addition, work is being undertaken with a time-delay variant of the RBF network to gather temporal information. In this way, simple behaviours, such as $y$-axis head rotation, can be extracted from previously-tracked image sequences.
Keywords: Face recognition, neural networks, RBF networks, receptive field functions, image sequences, invariance
Are the physical properties of the approaching object the sole determiner of its successful detection, or does the psychological state of the driver have a role? This research therefore attempts to explain motorcycle accidents within the framework of psychological models of visual search, selective attention and expectancy.
My work can be described as eclectic. My current research investigates how a driver through experience develops an efficient search strategy enabling them to emerge onto a main road. This research involves the measurement of driver eyemovements in a laboratory environment. Other ongoing research involves collaborative work with Sussex police investigating driver speed perception.
Keywords: Conspicuity, visual search, selective attention, Error modelling
Age-perception independent of identity: the questions that I have been addressing are concerned with trying to establish how an individual makes a decision abou the age of another and what cues they are using for that task. I have also tried to explore devleopmental changes in accuracy and cue usage.
The role of age-information in a facial memory: this research is concerned with age-invariance - that the same face can be recognised (or accommodated within a memory) even after transformations due to ageing have occurred over significantly long periods of time. The empirical questions that I have been addressing are concerned with establishing the boundaries of age-invariance, for example, when does recognition break down? How is the current percpetion accommodated within memory suitable for future recognition? What is the role of age-information in a memory of a face?
Keywords: face perception, event perception, face recognition, invariance, ageing, growth
The neural control of eye movements is perhaps the best understood brain function in mammals. The inputs are well known, the outputs are the positions of the eyes, and the psychophysics and neuroanatomy have been studied in detail. In addition the purpose of the movements seems clear.
All this is in contrast to the `vision` system, where the inputs are well known, but anatomical studies are less comprehensive, and there is much dispute as to the 'purpose' and little agreement on the 'outputs' of the system.
However, as well understood as the oculomotor system is, there appear to have been few attempts to model the whole system. The field is divided along arbitrary, perhaps inappropriate, lines. Sometimes the dividing lines represent interesting features of the system. An example is that most researchers study either fast or slow movements. Yet a single task will elicit both fast and slow movements, and the interaction may be of more interest than either type alone.
My intention is to make a neural network model of as much of the oculomotor system as possible. The model will be based upon the existing literature in neuroantomy and neurophysiology. I will test the model by repeating experiments from the psychophysical literature. Hopefully we may learn something about the extent of our understanding of the oculomotor system.
Keywords: eye movements, superior colliculus, vestibulo-ocular reflex, accessory optic system, neural modelling, primate
Copyright, © University of Sussex School of Cognitive and Computing Sciences, 1996,1997
13 May 1997 - Jonathan Howell, jonh@sussex.ac.uk
Back
to Vision Home Page
Back to COGS Home
Page