A hierarchical method for recognition of three-dimensional objects from single image
George Chi-Seng Lai, Rui J. P. de Figueiredo · 2002
This dissertation is the culmination of the research, development, and integration of three components of a 3-D object recognition system. The first component consists of a novel technique for improving the detection of edge points from an image. A common criterion for edge detection is to find the threshold for rejecting non-edge image points. Such a threshold can be determined from the statistical distribution of the non-edge points. The new technique provides a more accurate approximation of this distribution than previous methods. As a result, a more accurate threshold can be found to reject a given percentile of non-edge points. The second component is driven by a new method for extracting low contrast edges from images. The method first creates a rough segmentation of an edge using a clustering process that accounts for both edge direction and edge strength, then refines the clusters using the split and merge paradigm. The current method emphasizes the procedural ordering of clustering followed by detection to ensure sensitivity to low contrast edges, while reducing the effect of noise during edge detection. The third component consists of an efficient method for recognition of 3-D objects from a single image using quantitative and qualitative descriptions from topological and geometrical features. The recognition system builds on fundamental features such as edges and vertices and groups them into facets. To allow for recognition using a single image, the current method adopts the technique of invariant feature indexing for identifying the facets. However, the facets by themselves do not identify the object, but form a new hierarchy of features at which the recognition process can be accomplished with higher efficiency. The object is recognized by a novel and efficient method for indexing object hypotheses based on topological relationship between the facets.