Feature detection plays an important role in object detection and registration. For data registration, correspondences of different frames in the same scene should be found firstly. However, in many cases, there are not effective correspondences, which leads to incorrect registration. One of the effective solutions is to add marker manually in the scene. A method for detecting 2D marker (a black circle drew on paper for scene, hollow circular for point cloud) automatically in 3D point cloud (3D position information only) was proposed. First of all, add a 2D marker in real scene, and divide data set of 3D scene into segments by using region-growing segmentation. Then for each segment, detect 2D marker by extended RANSAC doing shape fitting. By this method, the 2D marker in 3D point cloud could be effectively detected without deforming or changing object. It provides simple and available features for the scene that is lack of features, laying a good foundation for next steps.