Incorporating Polarization in Stereo Vision-based 3-D Perception of Non-Lambertian Scenes

Surfaces with specular, non-Lambertian reflectance are common urban areas. Robot perception systems for applications in urban environments need to function effectively in the presence of such materials; however, both passive and active 3-D perception systems have difficulties with them. In this paper, we develop an approach using a stereo pair of polarization cameras to improve passive 3-D perception of specular surfaces. We use a commercial stereo camera pair with rotatable polarization filters in front, of each lens to capture images with multiple orientations of the polarization filter. From these images, we estimate the degree of linear polarization (DOLP) and the angle of polarization (AOP) at each pixel in at least one camera. The AOP constrains the corresponding surface normal in the scene to lie in the plane of the observed angle of polarization. We embody this constraint, an energy functional for a regufarization-based stereo vision algorithm. This paper describes the theory of polarization needed for this approach, describes the new stereo vision algorithm, presents results on synthetic and real images to evaluate performance.