Aiming at the problem that fuzzy C-means clustering algorithm is sensitive to initial value and easy to fall into local optimal when used for point cloud segmentation, resulting in unsatisfactory and unstable point cloud segmentation effect. An improved Wolf pack algorithm based on curvature constraint is proposed to optimize fuzzy C-mean clustering(IWPAFCM). In this algorithm, a good point set is introduced into the Wolf pack algorithm to initialize the population distribution. Then, adaptive step size is used to simplify parameter setting, balance optimization and convergence time. The interaction strategy was further used to enhance the wolves' internal communication and improve the wolves' ability of global optimization. Finally, the Gaussian perturbation mechanism was added to the Wolf to make it have the ability to jump out of the local optimal. The clustering center obtained by IWPA was used as the initial value of fuzzy clustering for iteration, and the accurate clustering center was obtained. On this basis, the distance between point clouds is defined based on the normal vector and curvature of point clouds and the traditional Euclidean distance is replaced to achieve the ideal point cloud segmentation effect. The Chair and Stool point cloud model and the measured point cloud model of mechanical parts and automobile covering parts in ModelNet40 open data set were taken as examples to verify the feasibility of the algorithm, and compared with FCM、FAFCM、WPAFCM and MACWPAFCM algorithms. The results show that for the four point cloud models, compared with the four comparison algorithms, the proposed algorithm has an average improvement of 0.4% to 11.95% in the clustering performance index VPC with high value as the best. In terms of fitness function value Jm、VPE和VXB clustering index, and fitness function value with low value as the optimal value, the average number of iterations is reduced by 0.2%-11.97%, 0.65%-7.35% and 0.3%-19.47% respectively, and the average number of iterations on the two ModelNet40 point cloud models is reduced by 8-20 times. The average number of iterations on the two measured point cloud models is reduced by 39-57 times, indicating that the proposed algorithm has fast convergence speed, few iterations, good clustering effect, higher clustering accuracy and better comprehensive performance.