This paper presents a statistical analysis of the Keypoint-based 4-point Congruent System (K4PCS) algorithm’s performance in global image registration. The evaluation is conducted on three test sets: Armadillo, Bunny, and Cat. We evaluate the precision of transformation matrices using root mean square error and the Frobenius norm, as well as real-time performance. Each registration is performed over 30 iterations to assume a normal distribution of results. Our results indicate that K4PCS consistently achieves satisfactory registration accuracy when aligning identical point clouds, regardless of their initial positions, and that the registration times, especially with downsampling, suggest the potential for real time performance. However, performance decreases when we introduce perturbations by adding Gaussian noise or by removing data points from the datasets. In the presence of added Gaussian noise, registration results maintain normality assumption up to 15% added noise. Computation times remain stable and close to real-time performance for each noise level. For partial-to-whole registration, error rate stays on the same level as for whole-to-whole registration. On the other hand, registration times vary depending on the number of removed points.