A new correspondenceless geometric algorithm for automatic inspection of filter components

We report on the design and validation of a method for correspondenceless estimation of transformation parameters. The problem we are tackling is connected with the verification of dimensional specifications of filter components in a manufacturing line. Our proposed method consists of acquiring range images of the same object's region from known, different viewpoints and use this information to calibrate camera parameters. Calibration errors can then be used to reason about actual physical dimensions and compare these with given specifications. We are investigating a correspondenceless method because the establishment of point to point correspondences between real images is a truly challenging task. Our proposed algorithm is based on geometric properties of image points which can be defined as invariants uniquely determined by the specific rigid transformation. The same invariants are also defined for a small subset of data, providing that the subset contains at least three data points. The algorithm is based on computing two sets of possible transformation parameters using different subsets of image data. The algorithm makes full use of rigid geometric constraints of image points synthesised into a single coordinate system and provides a closed form solution to all parameters of interest. We also implemented another well known correspondenceless calibration algorithm based on the scatter matrix. Experimental results have shown that the proposed algorithm is generally more accurate than the scatter matrix based algorithm and that it can be effectively used in the context of the described industrial inspection task.