We propose an improved point cloud registration method based on point-by-point forward feature point extraction to improve the efficiency and accuracy of point cloud registration. Firstly, the point-by-point forward method was used to quickly extract the point cloud feature points, significantly reducing the number of point clouds while retaining the characteristics of the point cloud model. Then, the improved KN-4PCS algorithm using normal vector constraints was coarsely registered to achieve the preliminary registration of the source point cloud and the target point cloud. Finally, the two-way Kd-tree optimized LM-ICP algorithm was used to complete the fine registration. In this paper, registration experiments were conducted on different point cloud data. In the registration experiment on Stanford University open point cloud data, the average error was reduced by about 70.2% compared with the SAC-IA+ICP algorithm, about 49.6% compared with the NDT+ICP algorithm, and the registration time was reduced by about 86.2% and 81.9%, respectively, while maintaining high accuracy and lower time consumption after introducing different degrees of Gaussian noise. In the point cloud registration experiment on real indoor objects, the average registration error was 0.0742 mm, and the average algorithm time was 0.572 s. The experimental results show that the proposed method can effectively improve the point cloud registration’s efficiency, accuracy, and robustness, thereby providing a solid foundation for indoor target recognition and pose estimation based on the point cloud.