Aiming at the problem of high computational complexity of feature descriptors and low registration accuracy, a point cloud registration algorithm based on the differences of region′s feature information in different scales was proposed. In the aspect of feature descriptor, the neighborhood spaces with different scales were selected for the key points. The normalized eigenvalue vector differences and normal vector angles between the scales were calculated. The descriptor of key point based on neighborhood scale differences was created. It is simple and time-saving. For the key points searching, a key point extraction method based on shape index was designed. The obtained key points have better representative ability. For searching the corresponding relationship, a double screening method based on Euclidean distance was proposed to find the correspondence set. The global optimal searching algorithm based on global distance was designed to find the transformation matrix between two point clouds. The experimental results show that the registration algorithm has good accuracy and robust noise robustness.