Line feature-based recognition using Hausdorff distance

A point-based (edge pixels) correlational method using the Hausdorff distance to determine if there is any model pattern in a given image was proposed by Huttenlocher et al. (1993). While this approach works well and it is computationally efficient in the presence of model translation in the image, it is significantly time consuming when the model has been rotated and scaled. We propose a line-feature based approach for model based recognition using the Hausdorff distance. This new approach reduces the problem of finding the rotation and scaling to the problem of finding two translations, therefore exploiting the efficiency of the Huttenlocher algorithm. The use of line features separates the rotation, scaling and translation so that each of them can be handled individually. The line features in the original domain are first transformed into a new 2D domain consisting of the orientation and the logarithm of the length of the line. In this way, rotation and scaling in the original domain correspond to a translation in the new domain and the Hausdorff point-based matching is used to find it. Next, the model is rotated and scaled using the result from first matching and second Hausdorff distance matching is performed to determine the model translation. The method performance and sensitivity to segmentation problems are characterized using and experimental protocol with simulated data. It was found that the algorithm performs well, degrading nicely as the segmentation problems increase. The algorithm was also tested with real images.