Abstract
While significant progress has been made in the computer vision field over the past decade, and machines capable of performing specialised visual inspection tasks are now being used in many industrial applications, the problem of recognising three-dimensional objects from two-dimensional imagery remains an area of ongoing research. Vision is undoubtedly our most important sense, and solutions to the problem of general three-dimensional machine vision must be found if the long term goal of autonomous robotic agents interacting naturally with humans in the real world is to be realised. In this work the problem of recognising polyhedra from two-dimensional images is investigated. The use of perceptual grouping and intermediate-level geometric features is considered, in particular the "edge-triple" feature. The edge-triple feature consists of three connected straight edges of an object, projecting to a triple of connected lines in the image, and can be used as a key feature, or indexing primitive, in model-based object recognition. The geometric constraints provided by matching such a configuration of image lines to an edge-triple are sufficient to uniquely determine the pose of the object. A probabilistic analysis of the edge-triple feature is performed, and a method for computing the probability densities of the angles formed by the projections of object edges under parallel projection is developed. These probabilities are used to prioritise the processing of potential model/scene feature matches produced by the hypothesis generation stage of a polyhedral object recognition system, substantially increasing the efficiency of the verification stage of the recognition while imposing negligible computational and storage penalties on the method. A new polyhedral object recognition system based on geometric hashing is implemented using edge-triple features. The method relies on extensive preprocessing of object models to encode invariant object data in a hash table. By performing as much of the object analysis as possible off-line, the efficiency of the actual recognition stage is maximised, at the expense of heavy demands on memory due to the large amount of data stored in the hash tables. However, the memory requirements of our edge-triple method are lower than those of conventional geometric hashing algorithms. Additionally, since our method employs lines and line groupings as key features rather than sets of interest points, our method is less susceptible to noise in the imaging and feature extraction stages than conventional geometric hashing. The validity of these assertions is demonstrated by extensive testing and evaluation of the method using both synthetic and real image data. It is demonstrated that the accuracy of pose estimates produced by the method is commensurate with theoretical predictions based on algorithm parameters and the expected errors in the image feature extraction. Since the projection from three-dimensional space to a single two-dimensional image necessarily involves a loss of information, the question of combining information from several images is addressed. A multi-view viewpoint consistency constraint is proposed, enabling the compatibility of recognition hypotheses from several viewpoints to be confirmed prior to the computationally expensive pose determination stage. The extra constraints provided by a multi-view analysis increase the reliability and robustness of the recognition system, while the consistency constraint helps to maintain the efficiency of the system. An active method to determine the complete three-dimensional structure of an edge-triple feature from two images is described. Finally, the limitations of the methods proposed and potential solutions to these shortcomings are discussed. Potential directions for future research are proposed.