Research On Feature Extraction And Recognition For 3D Object

The traditional algorithms of 3D object tracking, recognition, and pose estimation are based on point features for robotic manipulation and object grasping. Thereby, those methods usually fail with images with a lot of noise and texture, or low-texture scenes and objects common in man-made environments. In contrast to point features, geometrical features including line, contour, plane feature, are often abundant in man-made objects. Those features are more robust to texture, illumination, noise etc., and can handle the problem of partly sightlessness due to occlusion or the influence of illumination. The accuracy of 3D object tracking, recognition, pose estimation etc. can be improved effectively as well.This thesis systematically studies 3D feature extraction including line feature extraction and matching, contour feature extraction and matching, plane extraction as well as 3D object recognition based on geometrical features. The main work and researching achievements of this thesis can be summarized as follows:(1)Research on line extraction method. We use Canny edge detector to extract edges, and then the edges are split at points with high curvature. Finally, line segments are fitted into the spit edges by the least squares method and two segments are merged if their gap is small enough. In contrast to the traditional ones, this method is not so sensitive to the parameter setting of Canny detector.(2) Research on line matching method. We propose a line matching method based on local and global appearance, which is highly robust to the low-texture or repeated texture objects. Comparisons on real images show that the proposed method is more robust under the conditions of common image changes than the state-of-art methods.(3) Research on plane extraction method. This paper introduces a plane detection method using residual histogram of samples and J-Linkage without the prior knowledge of how many planes the object contains, which can use the matching set of points or lines as the input.(4) Research on contour extraction and matching method. A closed contour extracting method based on SD graph is studied and realized, which is robust to clutter and heavy light disturbance, and an AMTAR-based contour matching method is realized as well.(5) Research on 3D object recognition method. In order to improve the recognition capacity, a binocular vision based object recognition approach combining the depth and feature information is proposed.