Sensor planning and modeling for machine vision tasks

摘要

In this thesis, we present techniques to automatically synthesize desirable camera views of a known scene. Desirability of a camera view of a scene is represented in terms of a set of constraints. These constraints express whether certain scene features of interest are detectable or not in the resulting image. The feature detectability constraints that are chosen are fairly generic to vision tasks and require that the features are: (1) not occluded--the visibility constraint; (2) resolvable to a given specification--the resolution constraint; (3) in-focus-- the focus constraint; (4) within the field-of-view of the camera--the field-of-view constraint. An in-depth analysis of each of the above constraints results in the locus of viewpoints that satisfy each constraint separately. In this work, a viewpoint is an eight-dimensional quantity that consists of the three positional and two orientational degrees of freedom of camera placement, and three optical parameters of camera and lens setting. In order to determine globally admissible viewpoints, the loci of the individual constraints are combined by posing tile problem in a constrained optimization setting. Using existing optimization schemes, viewpoints that are globally admissible and locally optimal are obtained. In order to realize such a computed viewpoint in an actual sensor setup, the relationships mapping the planned parameters to the parameters that can be controlled, are obtained. This mapping is determined in the case of a sensor setup that consists of a camera in a hand-eye arrangement equipped with a lens that has zoom, focus and aperture control. The lens is modeled by a general thick lens model with non-coinciding pupils and principal points. The sensor planning and sensor modeling techniques that have been developed compose the MVP system. MVP is used in a robotic vision system that consists of a camera with a controllable lens mounted on a robot manipulator. The camera is positioned and its lens is set according to the results generated by MVP. Camera views taken from the computed viewpoints verify that the feature detectability constraints are indeed satisfied.