In order to improve the accuracy and efficiency of point cloud registration when there are large view changes, a point cloud registration method based on affine-invariant feature cloud purification and improved stochastic gradient descent is proposed in this paper. Firstly, the 2D feature matches with the ability to resist the change of view changes are obtained, and the point cloud purification method is designed to estimate the initial pose transformation of the point cloud based on the spatial topological relationship of the feature cloud. Then, on the basis of stochastic gradient descent method, a fast clustering nearest neighbor search strategy is designed to enhance the efficiency of searching for the corresponding points. The learning rate of the stochastic gradient descent is dynamically adjusted in probability to improve the global convergence. The experimental results show that the proposed point cloud registration method has a good adaptability to large viewing changes, and can effectively improve the accuracy and efficiency of registration.