To address the problem that the ICP algorithm relies heavily on initial values and may fall into local optimum solutions during the alignment process, an improved ICP algorithm incorporating geometric features is proposed. Using the feature that the part has rich geometric parameters, the curvature-based voxel sampling is first performed on the measured point cloud to preserve the geometric features as much as possible, and then the curvature difference and normal vector angle difference of the point cloud are introduced into the target error function of the algorithm for iterative calculation, and the final alignment results are obtained when the target error function reaches a set threshold. Experimental verification of point cloud alignment is carried out using standard parts with complex surfaces, and the results show that the improved ICP algorithm incorporating geometrical features converges faster and has lower errors than the ICP algorithm, and that the improved ICP algorithm incorporating geometrical features reduces the need for initial values and simplifies the point cloud alignment process while maintaining the accuracy of the alignment compared with the fast global registration plus ICP alignment algorithm. The improved ICP algorithm incorporating geometric features provides a strong support to facilitate accurate digital measurement and evaluation of parts, and has a technical reference value.