Color Constancy for Mobile Robots via the Graphics Rendering Equation

AAAI-10 PGAI Track Submission ID 54

Color Constancy for Mobile Robots via the Graphics Rendering Equation

Real-time object tracking is a basic, yet very important functionality for vision-based robots interacting with humans, indoors and outdoors. Among many descriptive properties of the objects, color is an easily visible, but highly illumination-susceptible property to track. Since an object’s true colors are often not perceived by the camera (due to varying illumination), a form of color constancy, the ability to interpret colors of objects as they are rather than how they are perceived, is useful for vision-based mobile robots. Traditionally, the problem of color constancy has been viewed as a mapping problem from perceived color to true color based on estimated illumination. In this paper, the problem is re-framed as mapping of known illumination and true color to perceived color. Using a generate-and-test methodology, we evaluate which illumination condition leads to a perceived color that most closely matches what the robot actually sees. To generate realistic views of an object under previously unseen conditions, we use lighting simulations established by the computer graphics community. One main contribution of this paper is determining the relationship between the vision community’s color constancy equation and the graphics rendering equation. We then apply such illumination simulations on multi-colored objects to achieve color constancy for vision-based mobile robots. The tracker is fully implemented on a Segway RMP for the task of following a person, and shows good real-time results in illumination-varying scenarios.

Duration : 0:1:46