An automatic point clouds registration method is proposed based on geometric information. This method can realize point cloud automatic registration without additional intensity and images information. Feature points of registration scans are extracted by feature distance histogram, and the initial matching sets of matching points are searched by K closet neighbor search in feature space. Relative height and the normal vector similarity measure are put forward, false match is eliminated combining with the root mean square distance and the initial conversion parameters are calculated. Points of low complexity are selected by Shannon entropy to participate in iterative closet point (ICP) accurate registration. Experimental results of real point cloud data show that, the proposed method can eliminate false matching of initial matching set, get better initial transform parameters, and improve the efficiency and precision of ICP algorithm.