A robust match filtering algorithm for use with repetitive patterns

Reliably matching feature points is an important part of many computer vision applications. This task is made harder when matching scenes containing repetitive patterns. The description of many feature points may be identical causing ambiguity in the matching results. This paper presents a filtering algorithm to remove erroneous matches caused by repetitive patterns. The proposed algorithm geometrically segments feature point locations into localized groups which are checked for consistency using correlation. A hierarchical approach is taken whereby neighboring groups are checked for consistency and collapsed into stronger ones. Finally a global model is calculated and used to ensure all cliques satisfy the scene geometry. The proposed method is generic and does not rely on specific feature detection algorithm. Experimental results demonstrate that the proposed method is superior to current state-of-the-art algorithms in accuracy and efficiency. The accuracy of matching repetitive patterns obtained from the proposed method is up to 99% compared to 96% obtained by previous state-of-the-art matching algorithms. The root mean squared residual matching error has been improved to 1.11 pixels compared to 4.09 obtained from current state-of-the-art image matching algorithms. The execution time of the method is competitive with most state-of-the-art image matching algorithms.