Recognition and Spatial Organization of Objects in Natural Scenes

Abstract

A computer vision model for recognizing objects in real world scenes and locating them in three-dimensional space is described. It is intended for applications which require navigation through an environment and interaction with objects in it. The system uses a recognition cone for feature extraction and preliminary recognition, a segmentation algorithm and a routine for constructing a three-dimensional world model. Several types of visual knowledge are incorporsted into the system including long-term object models, short-term object representations, and general routines for interpreting perspective shadows, highlights, occlusions, and texture gradients. The vision model is currently implemented in SIMULA and is being tested on near views of outdoor scenes. Results are presented for one scene of a house and yard. The design of a monocular motion parallax routine is given which will permit the system to integrate several views of one scene over time.