A registration algorithm matches two point clouds into one same coordinate system, which is widely used in object reconstruction and positioning of autonomous cars. The traditional point cloud registration algorithm such as an ICP cannot meet the requirements of initial value independence and real-time, furthermore, most improved ICP algorithms are based on the extraction of a single feature. Since PointNet++ is a deep learning model that can directly consume a disordered point cloud, we combine PointNet++ with ICP for a new registration method. Multiple features can be extracted by PointNet++, and these features are used as the basis for the registration. Then, the rotation and translation can be calculated by ICP algorithm. Experiments show that our registration method can achieve fast and robust registration.