Geometric feature distributions for shape representation and recognition

One of the fundamental problems in computer vision is the identification of objects from their shape. The research reported in this thesis is directed toward the development of a scheme for representing the shape of an object which allows it to be recognised both quickly and robustly across a wide range of viewing conditions. Given a shape described by a set of primitive elements, eg. straight line segments, the proposed scheme involves using a histogram to record the distribution of geometric features, eg. angle and distance, measured between pairs of primitives. This form of shape representation has a number advantages over previously proposed schemes. Foremost among these is the fact that it is able to produce local representations of shape, based on individual line segments. Recognition based on such representation is robust to the problems arising in cluttered scenes. Representations produced by the scheme are also invariant to certain object transformations, they degrade gracefully as the shape is fragmented and are strong enough to support discrimination between dissimilar objects. By treating the histogram recording a geometric feature distribution as a feature vector it is possible to match shapes using techniques from statistical pattern classification. This has the advantage that optimal matching accuracy can be achieved using processing which is both simple and uniform. The approach is therefore ideally suited to implementation in dedicated hardware. A detailed analysis is undertaken of the effect on recognition of changes in the description of a shape caused by fragmentation noise, scene clutter and sensor error. It is found that the properties of both the representation and matching components of the system combine to ensure that recognition is, in theory, unaffected by fragmentation noise, while it is maintained to very high levels of scene clutter. The factors which determine the effect of sensor error on the performance of the recognition system are fully analysed. The ability of the representational scheme to support object recognition is demonstrated in a number of different domains. The recognition of both 2D and 3D objects from a fixed viewpoint is demonstrated in conditions of severe fragmentation noise, occlusion and clutter. The scheme is then shown to extend straightforwardly to the representation of 3D shape. This is exploited to perform recognition and localisation of 3D objects from an arbitrary viewpoint, based on the matching of 3D scene and ,model shape descriptions. Finally, the use of the scheme within a multiple view-based approach to 3D object recognition is demonstrated.