Projective Invariance and Orientation Consensus for Extracting Boundary Configurations

Abstract

This work focusses on boundary detection of target objects by extracting configurations of straight lines. Hough transformation is used as basic voting technique for extracting the image lines. To find the appropriate peaks in the bin array of the Hough transform and thus estimating the line parameters a three-step procedure is involved. First, the Hough transform is constrained appropriately by the local orientations of grey level edges. Second, the global histogram of edge orientations is used to more or less emphasize certain parts in the Hough image. With these two modifications of the standard Hough transform we incorporate principles of orientation consensus and thus sharpen the peaks in the parameter space. Third, parallelism as a quasi-invariant of perspective projection is used to apply a sophisticated search strategy. Throughout the approach no specific object models but only basic principles of the imaging process are involved in extracting the boundary lines.