Abstract
Computer vision problems, such as tracking and robot navigation, tend to be solved using models of the objects of interest to the problem. These models are often either hard-coded, or learned in a supervised manner. In either case, an engineer is required to identify the visual information that is important to the task, which is both time consuming and problematic. Issues with these engineered systems relate to the ungrounded nature of the knowledge imparted by the engineer, where the systems have no meaning attached to the representations. This leads to systems that are brittle and are prone to failure when expected to act in environments not envisaged by the engineer. The work presented in this thesis removes the need for hard-coded or engineered models of either visual information representations or behaviour. This is achieved by developing novel approaches for learning from example, in both input (percept) and output (action) spaces. This approach leads to the development of novel feature tracking algorithms, and methods for robot control. Applying this approach to feature tracking, unsupervised learning is employed, in real time, to build appearance models of the target that represent the input space structure, and this structure is exploited to partition banks of computationally efficient, linear regression based target displacement estimators. This thesis presents the first application of regression based methods to the problem of simultaneously modeling and tracking a target object. The computationally efficient Linear Predictor (LP) tracker is investigated, along with methods for combining and weighting flocks of LP’s. The tracking algorithms developed operate with accuracy comparable to other state of the art online approaches and with a significant gain in computational efficiency. This is achieved as a result of two specific contributions. First, novel online approaches for the unsupervised learning of modes of target appearance that identify aspects of the target are introduced. Second, a general tracking framework is developed within which the identified aspects of the target are adaptively associated to subsets of a bank of LP trackers. This results in the partitioning of LP’s and the online creation of aspect specific LP flocks that facilitate tracking through significant appearance changes. Applying the approach to the percept action domain, unsupervised learning is employed to discover the structure of the action space, and this structure is used in the formation of meaningful perceptual categories, and to facilitate the use of localised input-output (percept-action) mappings. This approach provides a realisation of an embodied and embedded agent that organises its perceptual space and hence its cognitive process based on interactions with its environment. Central to the proposed approach is the technique of clustering an input-output exemplar set, based on output similarity, and using the resultant input exemplar groupings to characterise a perceptual category. All input exemplars that are coupled to a certain class of outputs form a category - the category of a given affordance, action or function. In this sense the formed perceptual categories have meaning and are grounded in the embodiment of the agent. The approach is shown to identify the relative importance of perceptual features and is able to solve percept-action tasks, defined only by demonstration, in previously unseen situations. Within this percept-action learning framework, two alternative approaches are developed. The first approach employs hierarchical output space clustering of point-to-point mappings, to achieve search efficiency and input and output space generalisation as well as a mechanism for identifying the important variance and invariance in the input space. The exemplar hierarchy provides, in a single structure, a mechanism for classifying previously unseen inputs and generating appropriate outputs. The second approach to a percept-action learning framework integrates the regression mappings used in the feature tracking domain, with the action space clustering and imitation learning techniques developed in the percept-action domain. These components are utilised within a novel percept-action data mining methodology, that is able to discover the visual entities that are important to a specific problem, and to map from these entities onto the action space. Applied to the robot control task, this approach allows for real-time generation of continuous action signals, without the use of any supervision or definition of representations or rules of behaviour.