Three-dimensional object description and recognition using feature vector techniques
Russell W. Taylor · 1988
A general method for identifying 3-dimensional objects from arbitrary viewing angles has been developed which involves novel fast matching techniques. A polyhedral sampling technique is used for generating an exhaustive set of library views, and a complementary set of worst case test views. Trials with these views are then used to create error maps, which show the distribution of viewing angles yielding incorrect classifications. Based on these error maps, the use of multiple views for object identification is examined. Next, a two level classification quality assessment is developed. The first level rejects objects that are not contained in the model database. The second identifies the likelihood of error for classifications of known objects. A number of different feature encoding techniques may be used for object representation with this scheme. Experiments have been conducted with Fourier descriptors and moment feature vectors, which have been computed from, respectively, contour imagery, and silhouette or range imagery. Image segmentation and the processing of three dimensional segments have been explored to extend the above system to work with partial object views. An algorithm has been developed for rapidly dividing surfaces in range imagery into planar regions. A split-and-merge segmentation approach is used, where the homogeneity criteria is based on a three parameter planar surface description technique. The importance of merge ordering is considered; in particular an effective ordering technique based on dynamic criteria relaxation. Many image processing operations for segment manipulation, normally carried out in image space, can be performed in moment space. A finite subset of the moment space describing a given image, is defined, on which many typical 2-dimensional geometric operations such as image plane rotation, and some 3-dimensional operations such as translation, are closed operations. An extended version of this set is introduced for which 3-dimensional rotation is also a closed operation. Techniques to perform moment space convolutions are presented. A new least squares procedure for polynomial surface fitting, is developed for application in moment space. The 3-dimensional normalization of image surfaces using these methods and a resulting general partial shape recognition scheme are considered.