A automatic registration algorithm for point cloud based on feature extraction is presented for threedimensional point cloud data obtained under different views. The algorithm extracts the feature points according to the variation of normal vector within various radius, and finds matching point pairs by using three geometric features of local point clouds comprehensively. The accurate matching point pairs by using distance restriction are obtained and the initial registration parameters are calculated. During the phase of accurate registration, the improved iterative closest point (ICP) should be used to finish the second mosaicking. The experimental results show that the proposed algorithm is more effective than traditional ICP in terms of run time and accuracy.