Orientation histogram-based matching for region tracking

Abstract

A region tracking technique with particular emphasis on rotation robustness is presented. It is based on region matching divided in two consecutive steps, gradient orientation histogram matching and template matching through normalised cross correlation (NCC). Given the orientation histograms of two image patches, a novel technique is used to estimate the rotation between them together with the similarity. This estimation enhances the performance and speeds-up the process of patch recognition. Fast computation of histograms using the integral histogram approach is exploited. Experiments show a high accuracy in the estimation of location and orientation.