ABSTRACT
The scale space analysis and corner detection scheme presented in the last chapter holds for curves with uniformly spaced points only. The scale space analysis does not involve convolution of the curve with Gaussian kernel. Corner detection is done without estimating curvature. In this chapter we propose another scale space analysis technique followed by corner detection. The procedure involves convolution of a digital curve with a smoothing kernel. Corner detection is done via curvature estimation. The procedure holds for curves with uniformly as well as non-uniformly spaced points. Rattarangsi and Chin [119] make scale space analysis using Gaussian kernel with varying window size. So the space requirement for the Gaussian filter coefficients is of the order of the square of data size. In this chapter we present an alterna-
tive approach to multiscale corner detection using iterative Gaussian smoothing with constant window size. As the window size is held constant, the space requirement for the Gaussian filter coefficients is finite and independent of data size. In the following chapter we present the iterative Gaussian smoothing process, show its convergence and the maximum allowable number of iterations that can be performed on a closed digital curve without wrap around effects. A scale space map showing the location of the maxima of absolute curvature is proposed. The map is shown to enjoy scale space property. An analysis of the scale space behavior of corner models is presented. Based on this analysis a tree organization is designed and corners are detected and located in a process of interpreting the tree. The space requirements and computational load is discussed and compared with [119]. Experimental results are presented to show the performance of the corner detector.
