Abstract:

Aiming at the problems of narrow convergence region and robustness in classical point cloud registration algorithms,a new algorithm for point cloud registration was proposed based on kernel density estimation.A new measure was proposed and used to evaluate the similarity between kernel density functions,
which provided a smooth bridge between the Kullback-Liebler divergence and the Euclidean distance.
The analytical form measure was derived under rigid constraints.The local maximum phenomenon caused by over-scaled parameter and the drifted maximum phenomenon caused by deficiency-scaled parameter were analyzed based on comparative experiments.And then,the variable-scaled BFGS-quasi Newton method was used to search the optimal parameters.Experimental results show that the method can
realize point cloud registration,extending the convergence region,meanwhile,improve the robustness under white noise interferences.

Key words: point cloud registration, kernel density estimation, measure function, BFGS-quasi Newton method

摘要：

针对传统点云配准算法收敛区间窄、鲁棒性差的难题，提出了一种基于核密度估计的点云配准算法。构建了一种能够实现Kullback-Liebler测度与欧氏测度之间平滑过渡的核密度分布相似性测度，推导了该测度在刚体约束下的解析表达式;通过对比实验分析了测度函数在大尺度参数下平滑但存在极值偏移，在小尺度参数下全局极值位置精确但存在局部极值的性能特点;提出采用尺度参数可变的BFGS拟牛顿算法进行点云配准参数的寻优求解。实验结果表明,该算法实现了点云数据的配准，拓展了算法收敛区间，同时提高了算法在白噪声干扰下的鲁棒性。



关键词: 点云配准, 核密度估计, 测度函数, BFGS拟牛顿法