Moving cast shadow detection based on region intrinsic image with entropy minimization

Abstract

Moving cast shadow detection is critical for outdoor video surveillance system because shadow points are often mis-classified as object points, which causes errors in segmentation and tracking. We propose a novel approach using intrinsic image combined with a model-based classifier to segment cast shadow. First, the moving object and its shadow are segmented roughly by background subtraction. Then, the color invariant areas are identified by intrinsic image based on illuminant invariance. We take a region comparison method instead of pixel by pixel to determine shadow area in these two steps. Meanwhile, a conclusion is drawn that using dynamic method to construct histogram will cause instability. Accordingly, a simple algorithm for finding minimum entropy is built when projecting ID intrinsic image. The proposed algorithm is demonstrated on a video sequence. The performance comparisons show that the proposed method can detect penumbra and provides highest performance compared with the conventional ones.