Real-time detection of moving cast shadows using foreground luminance statistics

Abstract

This paper presents a robust real-time method for detection of moving cast shadows which employs the assumption of higher interdependence of luminance values for the shadow pixels in larger regions compared to the object pixels. First, a fast modified image differencing technique is used to separate foreground pixels from the background. Next, for a moving window of fixed width scanning the foreground regions, a new measure called Modified Correlation is introduced. The new measure is determined by first computing the correlation between the luminance values of the moving window and luminance values of its neighbouring windows; this correlation is then divided by a robust-to-noise range measured based on the luminance values of the moving window. The modified correlation exhibits abrupt rising transitions as it enters the shadow region from the object region, transitions which can be used to separate object pixels from shadow pixels. Thus, the new method is very effective at suppressing moving cast shadows, while avoiding limiting structures, unrealistic assumptions, the need for a-priori knowledge, and manual selection of critical parameters. An average shadow detection rate of 85.4% and an average shadow discrimination rate of 99.5% over multiple different sequences, higher than those of available methods in the literature, confirm the efficacy of the method. The robustness of the method is examined under different lighting conditions, different target-environment combinations, and sequences with object-shadow occlusion. The proposed method is computationally efficient and suitable for real-time situations.

Simulating Shadow Interactions for Outdoor Augmented Reality With RGBD Data

“…In video surveillance, many video-based shadow detection algorithms have been proposed for detecting the shadows of moving cars or pedestrians [15]- [17]. For example, Andalibi et al [17] developed a fast shadow detection approach in which each video frame is compared to continuously updated background image. They differentiate shadow areas from the foreground by assuming that the shadow pixels are associated with background pixels through a nonlinear tone mapping.…”