Abstract
Driving error is a major factor in majority of all traffic accidents. To address this problem, artificial cognitive systems based on human behaviour and inference of cognitive processes need to be developed for drivers safety and assistance in mundane scenarios. This research work aims to develop, analyse and demonstrate novel methodologies that can be useful for the design of cognitive driver assistance systems. In this thesis we propose to use a Perception-Action (P-A) learning approach to cognitive systems building for modelling human behaviour. The P-A approach seeks to reduce the complexity implicit within conventional environment/action-planning approaches, by mapping actions directly onto the resulting perceptual transitions, mitigating the use of intermediate representation and significantly reducing training requirements. As a prime strategy, we have decided to use machine learning techniques for building an adaptive system which provides descriptions of the driver’s intentional behaviour using a psychological P-A model for human intentional modelling: Extended Control Model. Due to the (Highway Code) protocol based nature of the research problem, we propose to use rule-based machine learning algorithms for characterising driver intentions (proof-of-concept model evaluation). A development and evaluation dataset comes from an instrumented car, comprising perceptual and driver-action/control inputs. For in-situ realtime applications, it is important that P-A modelling learning is considered as an online problem. As such, we propose a novel methodology that utilizes a variational calculus approach to optimize an objective function defining system’s prediction error, thus enabling P-A mapping to be treated as an online learning problem via gradient descent using partial derivatives. The proposed learning structure performs top-down modulation of low-level perceptual confidences (confidence functions of low-level input features) via the Jacobian of the higher levels of a Perception-Action hierarchy. Symbolic manipulation of perceptual confidences is carried using fuzzy-logic reasoning. A superior performance is achieved by the proposed learning framework compared to P-A learning without the top-down modulation. The approach developed also permits novel forms of context-dependant multi-level P-A mapping, important within the context of an intelligent driver assistance system.