Study On Forward-looking Object Image Matching And Localization

The forward-looking image matching and localization is one of challenging problems. For the depth of each point changes largely, the relationship of the template image and scene image can not be described by simple homography transformation but the complicated projective transformatioa The imaging viewpoint is unknown, the time may be different, and even the template image may be generated from the downward-looking image, so lots of information is lost.When the 3d model of the object is known, the matching and localization algorithm is given for matching the 3d points in model and 2d points in image. The 3d points are maintained by hand, so the model of 3d points is formed. The 2d points are extracted automatically. Using the relationship of the projective transformation between 3d points and 2d points, the translation and the rotation angle can be computed. The algorithm can output the transformation parameters and the correspondence simultaneously.When the 3d model of object is unknown, the template is described by the image around the object, and the image matching and localization algorithm based on the similarity of feature point is given. The procedures of extracting and describing the feature point are analyzed. The rule of determining the correspondence point is defined. The similarity of feature point is used to find correspondence points, and then the false correspondence points are eliminated through the epipolar constraint The transform parameters of images are computed and the localization of object is obtained. Experimental results show that the algorithm is robust in the case of scale factor, rotation angle, partial occlusion, and even 3d rotation angle change.The above method only considers the epipolar constraint and does not make full of the location information of feature points. In order to maintain the location information of the feature points, the principle of relaxation labeling method is analyzed and the local invariant is incorporated into the relaxation procedure. The algorithm based on relaxation labeling method and the feature point similarity is proposed. Experimental results demonstrate that the algorithm can find more correspondence points relative to the methods based on the epipolar constraintIn order to maintain the location information of the feature points, the principle of belief propagation method is analyzed and the local invariant is incorporated into the message passing procedure. The algorithm based on belief propagation method and the feature point similarity is proposed. Experimental results demonstrate that the algorithm can find more correspondence points relative to the methods based on the epipolar constraintThe three above methods first determine the correspondence points and then locate the object, but the characteristic of forward-looking image matching and localization is mat it does not care the accurate correspondence points but the position, scale factor and rotation angle of the object in the scene image, the algorithm based on mean shift and vote is proposed. Using the position, orientation and the descriptor of feature point, the match set is formed in the way of one to many. For each match point, the position, scale factor and the rotation angle of the object in scene image can be computed, so the 4d vote space is formed. The densest point in the vote space is kept as the position, scale factor and rotation angle of the object Experimental results of the algorithm demonstrate both the robustness and efficacy of the overall approach on real images.