It is often needed to paste many marked points to realize auto-registration in the process of large parts three-dimensional (3D) measurement. Because of the randomness of artificially pasted marked points and noise factors, mismatching marked points often exist in auto-matching, which affect the stability of point-clouds auto-registration for repeated measurements. For this problem, a method is presented which uses random sample consensus (RANSAC) algorithm to remove the mismatching marked points based on the auto-matching of marked points. The method divides all matching marked points into inner points and outer points according to the selected target model and related criteria, calculates the current optimum target model parameters using the inner points and finally calculates the best parameters after a certain times of random sampling. It effectively removes the distance and noise mismatching marked points which are generated in the process of point-clouds auto-registration of large parts. Simulation experiment and registration examples demonstrate that the method is practicable and improves the stability of point-clouds auto-registration effectively.