Statistical Approach to Affine
Invariant Matching with Line Features 1
Frank Chee-Da Tsai
Robotics Research Laboratory,
Courant Institute of Mathematical Sciences,
New York University
715 Broadway, 12FL
New York, N.Y. 10003
One of the most important goals of computer vision research is object recognition. Currently, most object recognition algorithms assume reliable quality of image segmentation, which in practice is often not the case. This report examines the combination of the Hough Transform with a variation of Geometric Hashing as a technique for model-based object recognition in seriously degraded single intensity images.
There is recently much focus on the performance analysis of geometric hashing. However, to our knowledge, all of them are focusing on applying the paradigm to point features and show that the technique is sensitive to noise. There is as yet no exploration of line features. In this report, we use lines as the primitive features to compute the geometric invariants for fast indexing into the geometric hash table containing the pre-processed model information. In addition, we analytically determine the effect of perturbations of line parameters on the computed invariant for the case where models are allowed to undergo affine transformation.
We have implemented the system with a series of experiments on polygonal objects, which are modeled by lines. It shows that the technique is noise resistant and suitable in an environment containing many occlusions.
1The author would like to thank Prof. Jack Schwartz and Isidore Rigoutsos for generous discussion.