优化球查询算法的点云分割

Point Cloud Segmentation of Optimized Ball Query Algorithm

为丰富PointNet++网络局部特征的表征能力、加强细节特征的表达效果、提高点云模型的分割精度,针对PointNet++中多尺度和多分辨率算法都采用的点集重叠划分方法—球查询算法进行了研究。在PointNet++中,球查询算法随机选取球形邻域内的特征点提取局部特征,导致局部特征表达效果欠佳。为加强局部特征表征能力,引入K-近邻优化策略,将球邻域内的特征点按照与中心点的距离由近及远排序。在局部特征提取过程中,当球邻域内的点数超过需要的特征点时选取距中心点相对较近的一批点作为局部特征提取点;当球邻域中的点数少于需要的特征点时,选取距中心点最近的特征点复制多次,补齐特征表示。将优化的球查询算法应用于PointNet++分割网络,利用S3DIS和ShapeNetPart作为数据集验证算法的有效性。实验结果表明,优化的球查询算法丰富了网络的局部特征表征能力,强化了细节特征的表达效果,提高了分割精度。

In order to enrich the representation ability of the local features of the PointNet++network,strengthen the expression effect of the detailed features and improve the segmentation accuracy of point cloud models,we study the sphere query algorithm,which is the point set overlap division method used in the multi-scale and multi-resolution algorithms in PointNet++.In PointNet++,the ball query algorithm randomly selects feature points in the spherical neighborhood to extract local features,which leads to poor local feature expression.In order to strengthen the local feature representation ability,we introduce the K-nearest neighbor optimization strategy,and the feature points in the spherical neighborhood are sorted from near to far according to the distance from the center point.During local feature extraction,when the number of points in the neighborhood of the ball exceeds the required feature points,select a group of points relatively close to the center point as the local feature extraction points.When the number of points in the neighborhood of the ball is less than the required feature points,the feature points closest to the center point are selected and copied multiple times to complement the feature representation.The optimized ball query algorithm is applied to the PointNet++segmentation network,and S3 DIS and ShapeNetPart are used as a data set to verify the effectiveness of the algorithm.The experiment shows that the optimized ball query algorithm enriches the local feature representation ability of the network,strengthens the expression effect of detailed features,and improves the segmentation accuracy.