Recognition of partially occluded objects using an orthogonal complex AR model approach

The application of the orthogonal complex AR model approach has been extended for the recognition of partially occluded objects. The 2-D boundary of an object is extracted and is divided into a number of line segments by locating the turning points. Each line segment is resampled into a fixed number of data points and a complex AR model is used to represent each line segment. An orthogonal estimator is implemented to determine the correct model order and to estimate the corresponding AR model parameters. Each line segment is associated with a minimum feature vector denoting the estimated AR model parameters and similar line segments for different patterns have similar AR model parameters. Recognition of an occluded object based on the matching of line segments of two patterns is derived. An algorithm for finding the turning points has also been devised, which is quite robust to various object sizes and orientation, resolution and noise effect of the object image. Experimental results were obtained to show the effectiveness of the proposed approach for the recognition of objects which are partly covered or partially occluded, and the approach is robust to different object sizes and orientation.