|
Abstract : |
Researchers in robot vision have access to several excellent image processing packages ?e.g., Khoros, Vista, Susan, MIL, and XVision to name only a few ? as a base for any new vision software needed in most navigation and recognition tasks. Our work in automonous robot control and human-robot interaction, however, has demanded a new level of run-time ?exibility and performance: on-the-?y con?guration of visual routines that exploit up-to-the-second context from the task, image, and environment. The result is Gargoyle: an extendible, on-board, real-time vision software package that allows a robot to con?gure, parameterize, and execute image-processing pipelines at run-time. Each operator in a pipeline works at a level of resolution and over regions of interest that are computed by upstream operators or set by the robot according to task constraints. Pipeline con?gurations and operator parameters can be stored as a library of visual methods appropriate for di?erent sensing tasks and environmental conditions. Beyond this, a robot may reason about the current task and environmental constraints to construct novel visual routines that are too specialized to work under general conditions, but that are well-suited to the immediate environment and task. We use the RAP reactive plan-execution system to select and con?gure pre-compiled processing pipelines, and to modify them for speci?c constraints determined at run-time. 1, |
