Multi-level 3D-tracking of objects integrating velocity estimation based on optical flow and Kalman-filtering

This contribution describes an extensively tested approach to threedimensional tracking of a known polyhedral object at two levels of visual robot control, a position-based tracking level and a velocity-based tracking level. At the higher position-based level, an iterated extended Kalman-Filter is used to track a workpiece at a control-cycle of about two hundred milliseconds. This slower cycle operates on top of the video-rate tracking-loop of the velocity-based tracking level. While features such as vertices, edges, and ellipses extracted from gray-value images are used to adjust the 3D pose-estimate at the higher level of control, optical flow is used to initialise the velocity parameters of the upper level and to locally update poses at the lower level.