To solve the problems of low accuracy and slow calculating speed of the traditional point cloud registration methods for vehicle-mounted lidar, an improved FPFH-ICP registration combining fast point feature histograms (FPFH) initial matching with improved iterative closest point (ICP) accurate registration was proposed. Firstly, voxel grid and statistical-outlier-removal filter were used for preprocessing data before registration. Then, based on sample consensus initial alignment (SAC-IA), FPFH was used for initial registration to provide good pose information for accurate registration. Finally, a K-D tree was established, and a normal vector threshold was added to traditional ICP registration for accurate registration. In the experiments of four different scenarios, the root mean square error and registration time of the improved FPFH-ICP registration were reduced by 7.56% and 41.22%, respectively, compared with ICP registration, and by 30.28% and 18.95%, respectively, compared with point feature histograms (PFH) registration, suggesting that the improved FPFH-ICP registration can achieve accurate and efficient registration of point cloud data of vehicle-mounted lidar.