一种基于改进粒子群优化的点云配准算法

A Point Cloud Registration Algorithm Based on Improved Particle Swarm Optimization

摘要: 针对FPFH-ICP算法经验参数较多，难以通过人工调节获取最优配准参数，达到最优配准精度的问题，本文提出一种基于改进粒子群优化(Particle Swarm Optimization, PSO)的PSO-FPFH-ICP算法。算法以最小化配准后的均方根误差(Root Mean Square Error, RMSE)为目标，通过改进的粒子群优化算法不断对配准参数进行调节择优，最终获取限制区间内最佳的参数，从而提升算法的自动化程度和精确度。通过普林斯顿大学提供的3DMatch中的子数据集SUN3D，共四对不同场景点云进行六次重复实验，并与常用的FPFH-ICP、RANSAC-ICP以及4PCS-ICP三种点云配准算法进行比较。本文算法相较于FPFH-ICP平均RMSE降低了9.010 mm，相较于RANSAC-ICP降低了8.606 mm，相较于4PCS-ICP降低了7.322 mm，并避免了繁杂的人工调参过程，证明了本文算法具有高自动化和精准配准的特性。

Abstract: Aiming at the problem that the FPFH-ICP algorithm has many empirical parameters and it is diffi-cult to obtain the optimal registration parameters by manual adjustment to achieve the optimal registration accuracy, this paper proposes a PSO-FPFH-ICP algorithm based on improved particle swarm optimization (PSO). The algorithm aims to minimize the root mean square error (RMSE) af-ter registration. The improved particle swarm optimization algorithm is used to continuously adjust and select the registration parameters, and finally obtain the best parameters in the limited inter-val, so as to improve the automation and accuracy of the algorithm. Through the sub-data set SUN3D in 3DMatch provided by Princeton University, a total of four pairs of different scene point clouds were repeated six times, and compared with the commonly used FPFH-ICP, RANSAC-ICP and 4PCS-ICP point cloud registration algorithms. The average RMSE of the proposed algorithm is 9.010 mm lower than that of FPFH-ICP, 8.606 mm lower than that of RANSAC-ICP, 7.322 mm lower than that of 4PCS-ICP, and the complicated manual parameter adjustment process is avoided, which proves that the proposed algorithm has the characteristics of high automation and accurate regis-tration.