为实现点云数据的区域划分, 提出一种基于改进的粒子群优化与模糊C-均值聚类的混合算法(SPSO-FCM算法)。针对在点云聚类过程中易过早捕获局部极小值的问题, 算法首先用改进的粒子群算法——社会粒子群优化算法, 对种群进行初始化, 通过为每一个粒子设置不同的跟随阈值, 来维护种群中个体多样性, 加深对种群全局搜索的程度, 避免陷入局部极小值; 随后, 设置种群中每个粒子当前最优位置和初始种群的最优位置, 更新自由粒子的位置和跟随粒子的速度和位置; 最后, 采用模糊C-均值聚类算法求解隶属度矩阵, 确定适应值函数, 更新所有粒子的最优位置, 并判断粒子和种群的位置优越性, 得到准确的聚类中心, 实现对点云数据的区域划分。以曲面复杂度不一致的点云模型为例对算法进行验证, 探讨SPSO-FCM聚类算法的可行性, 并与FCM聚类算法、遗传FCM聚类算法进行比对。实验结果显示, SPSO-FCM聚类算法较其它两种算法, 收敛速度快, 迭代次数少, 聚类准确, 边界区域分割清晰, 特别是对型面复杂、点云数据较多的机械零部件点云数据进行分割时, 能得到更好的分割结果。

To realize region segmentation of point cloud data, a kind of mixed algorithm (SPSO-FCM algorithm) based on improved particle swarm optimization and fuzzy-C means clustering was introduced. Aimed at local minimum easily to be captured untimely in point cloud clustering process, improved particle swarm optimization algorithm-social particle swarm optimization algorithm was used firstly to initialize population in the algorithm. By setting different follow thresholds for each particle,variety of individual in population was maintained and the global search degree of population was enhanced to avoid falling into the local minimum. Then the current optimal position of each particle in population and optimal position of initial population were set to update position of free particle and speed and position of following particle. Finally, fuzzy C-means clustering algorithm was adopted to solve membership matrix and determine fitness function. On the basis of above, optimal position of all particles were updated and position superiority of particle and population were judged to gain correct clustering center and to realize region segmentation of point cloud data. Took point cloud model with inconsistent surface complexity as example to verify algorithm and discuss feasibility of SPSO-FCM clustering algorithm and compare with FCM clustering algorithm and genetic FCM clustering algorithm. Experimental result shows that compared with other 2 algorithms, SPSO-FCM clustering algorithm has quicker convergence rate and less iteration with more correct clustering and clearer boundary region segmentation, and especially for point cloud data segmentation of mechanical components and parts with complex molded surface and numerous point cloud data, it can get better segmentation result.