| A model-based computer vision system is designed for multiple object 3-D motion interpretation. For given image sequences, the system recognizes 3-D rigid objects in the scene, determines their 3-D positions and orientations, and calculates their 3-D translations and rotations. The system achieves this goal by applying a set of consistency constraints, such as the viewpoint consistency constraint. The processing hierarchy of the system consists of three stages: feature extraction, model hypothesis generation and verification, and motion estimation. The system has the following new features: a control structure which employs both bottom up and top down processing strategies, a structure-based indexing technique for efficient object model selection, a viewpoint estimation method to obtain the correspondences between the image features and the 3-D object model features, and a multiple camera information fusion method. Sample applications are shown. |
