Model based object recognition by robust information fusion

Given a set of 3D model features and their 2D image, model based object recognition determines the correspondences between those features and hence computes the pose of the object. To achieve good recognition results, a novel approach based on robust information fusion is put forward in this paper. In this algorithm, the property of probabilistic peaking effect is employed to generate sets of hypothesized matches between model and image points. The correct hypotheses are obtained by searching for clusters among projections of predefined 3D reference points using the pose implied by each hypothesis. To assure the robustness of clustering, a new data fusion technique that is based on the nonparametric mode search method, mean shift, is proposed. The uncertainty information of the hypotheses is also incorporated into the fusion process to adaptively determine the bandwidth of the mean shift procedure. Experimental results demonstrating the satisfactory performance of this algorithm are presented.