Use this URL to cite or link to this record in EThOS:
Title: Learning object behaviour models
Author: Johnson, Neil
ISNI:       0000 0004 2696 2910
Awarding Body: University of Leeds
Current Institution: University of Leeds
Date of Award: 1998
Availability of Full Text:
Access from EThOS:
Access from Institution:
The human visual system is capable of interpreting a remarkable variety of often subtle, learnt, characteristic behaviours. For instance we can determine the gender of a distant walking figure from their gait, interpret a facial expression as that of surprise, or identify suspicious behaviour in the movements of an individual within a car-park. Machine vision systems wishing to exploit such behavioural knowledge have been limited by the inaccuracies inherent in hand-crafted models and the absence of a unified framework for the perception of powerful behaviour models. The research described in this thesis attempts to address these limitations, using a statistical modelling approach to provide a framework in which detailed behavioural knowledge is acquired from the observation of long image sequences. The core of the behaviour modelling framework is an optimised sample-set representation of the probability density in a behaviour space defined by a novel temporal pattern formation strategy. This representation of behaviour is both concise and accurate and facilitates the recognition of actions or events and the assessment of behaviour typicality. The inclusion of generative capabilities is achieved via the addition of a learnt stochastic process model, thus facilitating the generation of predictions and realistic sample behaviours. Experimental results demonstrate the acquisition of behaviour models and suggest a variety of possible applications, including automated visual surveillance, object tracking, gesture recognition, and the generation of realistic object behaviours within animations, virtual worlds, and computer generated film sequences. The utility of the behaviour modelling framework is further extended through the modelling of object interaction. Two separate approaches are presented, and a technique is developed which, using learnt models of joint behaviour together with a stochastic tracking algorithm, can be used to equip a virtual object with the ability to interact in a natural way. Experimental results demonstrate the simulation of a plausible virtual partner during interaction between a user and the machine.
Supervisor: Hogg, D. C. Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available