Task specification languages for uncalibrated visual servoing

摘要

In the field of vision-based robotic control, there is a large and growing body of work which investigates how to perform tasks based on direct image feedback. Work in this area, often called image-based visual servoing, has successfully demonstrated a variety of interesting and useful tasks. Little attention, however, has been paid to what tasks the image-based servoing framework makes possible. Answering the question “What tasks are performable with an uncalibrated stereo rig?” is the underlying motivation for this thesis. In particular, a formalism is presented that makes the above question precise and then provides a complete answer for a number of different possible meanings of “uncalibrated.” From these results this work builds a complete enumeration of distinguishable feature configurations for the considered sets of uncalibrated cameras. Such enumerations, in turn, lead to concise and provably complete languages for specifying tasks. These languages offer a small number of primitive skills and composition operators that together span the set of all tasks specifiable with image-based constraints. Applications of these formal results are also considered and tested. First, each of the primitive skills is shown to have an image-based constraint that is satisfied exactly when the underlying 3d geometric constraint is satisfied. The composition operators then provide a means for automatically generating image-based constraints for any specifiable task. Secondly, the full enumeration of distinct feature configurations allows a process to monitor whether or not enough feature points are available for a desired task specification. Example task specifications and results of their implementation are presented.