Object Recognition through Template Matching Using an Adaptive and Robust Hausdorff Distance

In computer vision, the partial Hausdorff distances (PHDs) are used to compare images but with strong limitations of using fixed fractions. In this paper, we formulate an adaptive and robust Hausdorff distance (ARHD) with nonparametric and robust statistical methods. The new distance is estimated using the empirical distribution of the distance variable based on the distance map of the template’s edge map, and this makes full use of the information associated with the edge distribution structure of templates. The best fraction is determined by adaptively adjusting at two directions along the distance curve and the statistical test using linear regression. The distance threshold is also derived from the same empirical distribution function estimated from the template. Therefore, it is not sensitive to the initial fraction values compared with conventional PHDs. The experiments using aerial images show that ARHD has good performance in matching templates in heavily blurred and complex backgrounds with pose change, scale change, geometric distortion and partial occlusion.