Abstract: A pavement crack detection algorithm was proposed on the basis of directional features and gravitational model to overcome the shortcomings of the existing global methods. According to the extension and local transition of cracks, the endpoints of cracks were firstly searched. The direction factors and directional derivative discrimination criterion were designed, which reflect the linear features of cracks. Crack discrimination was localized in the possible neighborhood to prevent effectively the introduction of redundant information. Based on the human vision characteristics and physics theory, a crack gravitational model was designed to improve the robustness of crack connection under the noise environment. By combining color distance, a multiple criteria discriminant function of crack extension and connection was designed to detect crack pixel area. The experimental results indicated the validity of the proposed method.