摘要
针对传统点云配准方法易受到噪点、离群值和重叠度的影响,造成配准精度低和效率低等缺点,提出了一种利用信息熵改进的欧式距离聚类中心点的方法来完成点云配准。首先对两片点云进行体素格网下采样,加快后续处理效率,不同于欧式聚类直接利用距离聚类,先计算点的特征值,根据特征向量求得的信息熵,利用特征向量来选取聚类,再提取出各类别的中心关键点,后使用KD-tree算法进行点对的搜索和对应,结合对应点对的位置信息估计出初始变换矩阵,作为精配准的输入矩阵,为后续精配准提供良好的初始位姿;最后采用双向KD-tree改进的点到面ICP算法进行精确配准。选用了长约300 m的道路点云数据进行实验,与四种方法在重叠度为10%时进行比较,结果表明算法的RMSE为0.074 m,总体配准过程消耗时长为30.256 s,比四种算法的配准精度和效率更高。
A method of Euclidean distance clustering of centroids using improved information entropy is proposed to complete the point cloud alignment,for the traditional point cloud alignment method is susceptible to noise,outliers and overlap,and solves the shortcomings such as causing low alignment accuracy and low efficiency.First of all,voxel grid down sampling is performed on the two point clouds to accelerate the efficiency of subsequent processing.Different from Euclidean clustering directly using distance clustering,this method computes the feature values of points.By calculating the information entropy based on the feature vectors,a feature tensor is employed for cluster selection.Subsequently,key points representing each cluster are extracted,and the KD-tree algorithm is employed for point pair searching and correspondence.Utilizing the positional information of corresponding point pairs,an initial transformation matrix is estimated,serving as input for precise registration and providing a favorable initial pose for subsequent refinement.Finally,a bidirectional KD-tree-enhanced point-to-plane ICP algorithm is employed for accurate registration.A road point cloud data with a length of about 300 m is selected for the experiment,and compared with the four methods at an overlap of 10%,the results show that the RMSE of the algorithm is 0.074 m and the overall time consumed by the alignment process is 30.256 seconds,which is higher than the four algorithms in terms of accuracy and efficiency of the alignment.
作者
喻俊楠
吴学群
赵辉友
YU Jun-nan;WU Xue-qun;ZHAO Hui-you(College of Land and Resources Engineering,Kunming University of Science and Technology,Kunming 650093,China)
出处
《激光与红外》
CAS
CSCD
北大核心
2024年第10期1541-1546,共6页
Laser & Infrared
基金
国家自然科学基金项目(No.41961053,No.41961039)资助。
关键词
点云配准
欧式聚类
信息熵
特征向量
中心关键点
点到面ICP
point cloud registration
Euclidean clustering
information entropy
feature vector
central key point
pointto-surface ICP
