Reconstruction and recognition of 3D objects from occluding contours and silhouettes
Chengchun Shien · 1987
Recognizing 3D objects from 2D images has been the central theme of research in computer vision. This dissertation presents a new approach to object recognition by using the information contained in the occluding contours of 3D objects. Two important issues related to object recognition are model construction and object matching. Due to their compact structure, quadtrees and octrees are chosen to describe 2D images and the 3D structure of objects, respectively. The octree of each model is generated from the quadtrees of the model using a technique known as volume intersection. This technique allows one to reconstruct both volumetric and surface information (including feature points) of an object from its multiple views. As a consequence, only the 3D feature points and the principal quadtrees of each model need to be stored in the library. Two techniques are developed for recognizing a 3D object for the cases where the observed object is viewed from multiple viewpoints and from a single viewpoint, respectively. In the multiple-views case, recognition is achieved by a two-stage matching, namely a coarse matching and a fine matching, based on the volumetric information reconstructed from the given multiple silhouettes. The coarse matching is applied to quadtrees and the fine matching to octrees. The octree of a model need not to be generated unless it is determined as a likely match in the coarse matching phase. This gives rise to a significant improvement in computational efficiency. In the single-view case, feature points are extracted to guide the recognition process. Four-point correspondences between the 2D and 3D feature points are hypothesized, and then verified by applying a variety of constraints to their associated viewing parameters. The problem of solving the viewing parameters for each hypothesized correspondence is formulated as a set of linear equations to ensure the efficiency. The result of the hypothesis and verification process is further verified by a 2D contour matching. This approach allows for a method of handling both planar and curved objects in a uniform manner. Furthermore, it provides a solution to the recognition of multiple objects with occlusion as well, since only four points are required for the initial estimation, and occluding contours are used for final recognition.