Three dimensional object recognition using linear features and oriented model points (hough)
Teresa M. Silberberg · 1985
This dissertation addresses the problem of determining the orientation and position of a three dimensional object in an image containing the projection of that object. Using linear features or oriented points from the model, a generalized Hough procedure is applied in order to detect global consistencies in the image data. First, we describe an iterative Hough procedure in which straight line segments in the image are matched by finding the parameters of a viewing transformation of a three dimensional model consisting of line segments. Assuming the scale of the object is known, there are three orientation and two translation parameters to be estimated. Initially a sparse, regular subset of parameters and transformations is evaluated for goodness-of-fit; then the procedure is repeated by successively subdividing the parameter space near current best estimates or peaks. The algorithm is demonstrated with models composed of linear features arising from surface markings and surface discontinuities. We next show that imposing retrictions on the scene yields a more efficient and more robust recognition algorithm. This new, two stage algorithm matches oriented features of an object model in the restricted scene to their perspective images in order to compute an initial estimate of the free rotational and translational parameters. This estimate is then refined by applying a top-down, least squares procedure to correspondences between the model and image data. An analysis of the effects of inaccuracies in extraction of the image features and measurement of camera and scene parameters on the matching procedure is provided. A method of recognizing objects in an image containing multiple objects is developed, and the use of additional data from the original image is discussed.