Realtime Estimation of Illumination Images Using Illumination Eigenspace

An illumination irnagc, which is a part of intrinsic images, represents the effect of a lighting condition of thc sccnc. To propcrly handlc illumination cffccts such as cast-shadows in the input image, image manipulation using thc illumination imagc is only natural, since it describes variation of lighting effects from a rcflcctancc imagc which can bc considered as an image under the standard illumination. We have shown in prcvious work (121 that illumination cffccts arc rcasonably factored out from the input images by using illumination imagcs. To apply this mcthod as a prcprocessing stage to a video surveillance system, realtime cstimation of illumination imagcs is rcquircd. Unfortunately, the cost of estimation of illumination images in rcaltimc is computationally high. In addition, it is necessary to synthesize background images before deriving illumination irnagcs whcn thc sccnc contains dynamic objects. In this paper, we illustrate our approach to modcling illumination irnagcs with principal component analysis (PCA) to directly estimate illumination imagcs from input imagcs which contain moving objects in the scene. We propose this framework presupposing that thc camcra is fixcd and thc sccnc is observed under several lighting conditions.