Recognising Three-Dimensional Objects using Parameterized Volumetric Models

Centre for Intelligent Systems and their ApplicationsThis thesis addressed the problem of recognizing 3-D objects, using shape information extracted from range images, and parameterized volumetric models. The domains of the geometric shapes explored is that of complex curved objects with articulated parts, and a great deal of similarity between some of the parts. These objects are exemplified by animal shapes, however the general characteristics and complexity of these shapes are present in a wide range of other natural and man-made objects. In model-based object recognition three main issues constrain the design of a complete solution: representation, feature extraction, and interpretation. this thesis develops an integrated approach that addresses these three issues in the context of the above mentioned domain of objects. For representation I propose a composite description using globally deformable superquadratics and a set of volumetric primitives called geons: this description is shown to have representational and discriminative properties suitable for recognition. Feature extraction comprises a segmentation process which develops a method to extract a parts-based description of the objects as assemblies of defoemable superquadratics. Discontinuity points detected from the images are linked using 'active contour' minimization technique, and deformable superquadratic models are fitted to the resulting regions afterwards. Interpretation is split into three components: classification of parts, matching, and pose estimation. A Radical Basis Function [RBF] classifier algoritm is presented in order to classify the superquadratics shapes derived from the segmentation into one of twelve geon classes. The matching component is decomposed into two stages: first, an indexing scheme which makes effective use of the output of the [RBF] classifier in order to direct the search to the models which contain the parts identified. this makes the search more efficient, and with a model library that is organised in a meaningful and robust way, permits growth without compromising performance. Second, a method is proposed where the hypotheses picked from the index are searched using an Interpretation Tree algorithm combined with a quality measure to evaluate the bindings and the final valid hypotheses based on Possibility Theory, or Theory of Fuzzy Sets. The valid hypotheses ranked by the matching process are then passed to the pose estimation module. This module uses a Kalman Filter technique that includes the constraints on the articulations as perfect measurements, and as such provides a robust and generic way to estimate pose in object domains such as the one approached here. These techniques are then combined to produce an integrated approach to the object recognition task. The thesis develops such an integrated approach, and evaluates its perfomance inthe sample domain. Future extensions of each technique and the overall integration strategy are discussed