Monocular 3D vision for a robot assembly environment

摘要

A procedure for updating estimates of an object's pose using information from one or more monocular images is presented. Features in monocular images are assigned correspondence with modeled 3D features based on estimated object position. An improved position estimate is computed based on feature correspondence. The method accommodates partial occlusion or contact among objects. Features need not appear in multiple views to be used for estimation. Results from this system are presented which demonstrate the location of multiple objects within approximately 0.1 in. in translation and 2° in rotation