POLARIZATION CAMERA FOR COMPUTER VISION WITH A BEAM SPLITTER

A fully automated system that utilizes two CCD cameras and a polarizin g beam splitter to create a polarization camera capable of sensing the polarization of reflected light from objects at pixel resolution is p resented. The physical dimensions of the polarization of light beyond that of intensity carry extra information from a scene that can provid e a richer set of descriptive physical constraints for the understandi ng of images. It has been shown that polarization cues can be used to perform dielectric and metal material identification and specular- and diffuse-reflection component analysis, as well as complex image segme ntations that would be immensely more complicated or even infeasible w ith the use of intensity and color alone. A polarizing-plate beam spli tter is placed in front of two CCD cameras so that light beams reflect ed from and transmitted through the beam splitter are each incident up on a separate camera. The polarization state of the reflected and the transmitted beams are linearly independent in terms of two orthogonal- polarization components, and these components are resolved in real tim e from the simple solution of two simultaneous linear equations. The p olarizing-plate beam splitter allows for the simultaneous measurement of two orthogonal-polarization components over fairly wide field views suitable for vision and robotics. A polarization contrast image can b e produced at 15 Hz. Two sets of orthogonal-polarization component pai rs can be resolved by electronically switching a twisted nematic liqui d crystal placed in front of the beam splitter, permitting the real-ti me measurement of partial-linear-polarization images at 7.5 Hz. A sche me for mapping states of partial linear polarization into hue, saturat ion, and intensity, which is a very suitable representation for a pola rization image, is illustrated. The unique vision-understanding capabi lities of this polarization camera system are demonstrated with experi mental results showing polarization-based dielectric and metal materia l classification, shape constraints from reflected polarization, and s pecular-reflection and occluding-contour segmentations in a fairly com plex scene.