Recognition of overlapping objects.

Abstract

This thesis describes the design and implementation of a model-based vision system for the recognition of partially occluded objects. The system can detect occlusion. It can also recognize and locate the objects forming the occlusion scene. The system has two modes of operation: a learning mode and a recognition mode. In the learning mode the objects to be recognized are analyzed, their boundaries processed and a model representing each object is created and added to a list of object models. In the recognition mode the system analyzes the image which may contain several objects, the objects may occlude each other. The system processes the image, detects whether there is occlusion and then identifies and locates the objects present in the image. The objects' models are based on the parameters of an auto-regressive (AR) filter. In the recognition mode the system first generates the scene sub-parts, then it goes through the three phases of the recognition: namely hypotheses generation, pruning and verification. (Abstract shortened by UMI.)