A two level model-based approach for object recognition
BEHROUZ PEIKARI, Ding-Chung Liu · 1988
The application of computer vision for automation and manufacturing process has received considerable attention in recent years, especially with the growing interest in robotics. To meet the requirements of speed, accuracy, and flexibility for robot vision, model-based approach is the most promising method among many techniques developed. The model-based approach is based on a predefined model and a matching algorithm associated with the properties of the model. To construct the model in a model-based approach, one of the following three methods is generally used: the global feature method, the structural feature method, and relational graph method. The matching algorithm is then developed in accordance with the characteristics of models. In addition to the model-based approach, an approach which uses a two level processes to handle a large set of objects is presented in this dissertation. In the first level, one of the subsets of the models is designated as the possible matching set by identifying the structures of the unknown object with a set of rules. Then, in the second level the proposed approach combines the constraint verification method and evaluates a distance criterion function to match the unknown object with the subset of the models for recognition as well as locating the position and orientation of the identified object. In order to achieve better efficiency in processing time, a parallel implementation of the matching is also employed. The proposed approach is simulated with 16 different images of objects on a Micro-Vax serial computer and a SEQUENT BALANCE 21000 multiprocessor computer to evaluate the performance of the proposed algorithm. The results show that high recognition rate can be achieved for the experimental sample images in controlled environment. The parallel implementation can efficiently speed up the processing for the matching task of recognition process.