Predictive computer vision for robotic assembly

Abstract

A description of the Workspace Prediction and Fast Matching System (WPFM System) is given. The purpose of the system is to exploit the special constraints on the identity and placement of objects in a typical industrial robot workcell to perform rapid analysis of stereo vision data derived from such a scene. The a-priori knowledge of the positions and identities of some objects can allow some stereo data to be quickly matched to model features, and subtracted out of the scene data, leaving unknown data to be dealt with by a more comprehensive (and more computationally expensive) vision matcher. A workspace prediction graph (WP Graph) for a given scene can be created using the Robmod CSG solid modelling system (the WP System). The 3D model features in the WP Graph can then be matched against 3D stereo data derived from the equivalent real scene, using the Fast a priori Matching System (FAPM System). Matching is currently carried out only with edge features. The stereo data used in the matching process is derived from the real scene by a low level stereo vision capability (the GDB System). Curved objects can be handled by the WPFM System, and these are represented in the WP Graph by polyhedral approximations, which are then matched to stereo data.