Stepping out of the Shadows Using Polarized Light

Hank Hogan

Researchers at the University of Pennsylvania in Philadelphia have demonstrated a machine vision technique that untangles multiple overlapping shadows within a scene, improving a camera’s ability to distinguish objects. It works by capturing polarization information instead of relying on the traditional parameters of intensity and hue.

A street scene acquired using data based solely on light intensity (left) does not provide as much contrast as an image acquired using information based on the degrees of polarization sensed at each pixel. The yellow circles indicate the same sewer drain in each scene. Courtesy of the Optical Society of America.

Principal investigator Nader Engheta, a professor of electrical and systems engineering, noted that polarization historically has not been widely used in this application, partly because it is effectively invisible. “Human eyes and ordinary cameras are practically polarization-blind,” he said.

Inspired by animals that can see polarization, the investigators looked into the use of the phenomenon to distinguish one shadow from the next. Shadow identification and segmentation are difficult for typical machine vision cameras and interfere with their ability to recognize and track objects.

The researchers used an Olympus digital camera with 4-megapixel resolution, 10-bit pixel depth and the ability to output unaltered sensor data. In front of the camera, they placed a polarizer, rotating it by 45° or 90° or leaving it as is. For various scenes, they captured each polarization angle, processing them on a pixel-by-pixel basis to create a degree-of-polarization image. From this, they could separate shadows created by various light sources, which would have been almost impossible with conventional imaging.

In one outdoor scene, for example, the degree-of-polarization image revealed a sewage drain cover and shadow patterns. In the same setting, using the traditional intensity-only image, the cover was concealed in shadow.

Behind this technique lies the fact that nonpolarized light is at least partially polarized when reflected from most flat surfaces. Therefore, diffuse sunlight — when reflected off windows or a building — will be partially polarized. As for indoor applications, some fluorescent lights and LCD monitors directly produce polarized light.

“It is very rare to find a scene without any polarized light, be it indoors or outdoors,” said postdoctoral researcher Shih-Schön Lin.

Knowing the degree of polarization enables at least partial determination of the sources of various intersecting shadows, thus permitting them to be separated. That, in turn, improves the contrast of machine vision.

Lin noted that the computational burden of the technique is not too great and that the processing can be performed in parallel. The group has demonstrated near-real-time extraction of some of the polarization information and is working on several methods of acquiring various polarization images — automating the process, for example, by using a mechanical filter rotator or a liquid-crystal polarizer sheet.

Optics Express, Aug. 7, 2006, pp. 7099-7108.