This paper proposes a multiple features fusion method for moving shadows removal mainly based on optical reflection invariant. It assumed that the optical reflection of object surface changed little when casted by shadows. Two techniques are designed for reflection-invariant estimation. One is homomorphic filtering that obtains reflection component from input image directly; the other technique is the proposed maximum channel prior. Based on homomorphic filtering, scale-invariant local ternary pattern is utilized to discriminate shadows from the foreground mask. Based on maximum channel prior, an approximate illumination invariant is designed by local neighboring information calculation. These techniques are utilized for weak or strong shadow detection. For weak shadows, multiple descriptors including color constancy, texture consistency, and reflection invariance at the pixel-level or region-level are combined based on shadow assumptions. For strong shadows, a coarse-torefine strategy is proposed using maximum channel prior technique. Experimental results on standard datasets consisting of 12 video sequences illustrate the proposed method's effectiveness compared with some state-ofthe-art approaches.