基于PointNet++的逆密度点云识别与分割算法

Inverse density point cloud recognition and segmentation algorithm based on PointNet++

摘要:

为了提高点云处理精度，针对PointNet++对不均匀分布的点云数据特征提取不完整以及忽略了部分点云特征导致分类与分割结果不佳等问题，对算法PointNet++进行了研究，提出了基于PointNet++的融合密度信息的逆密度点云识别与分割算法D-PointNet++；利用点云密度计算出每个点的采样概率，根据采样概率使用多项分布进行点云采样；通过自适应缩放分组半径进行点云分组；采用多种池化方法混合提取点云特征并利用多头注意力机制计算出多种特征的权重，并加权聚合得到点云的全局特征；实验结果表明，相较于多种参评算法，D-PointNet++在点云分类准确率、分割精度上均有显著提升。

Abstract:

In order to improve the processing accuracy of point cloud, PointNet++ algorithm was studied to solve the problems of incomplete extraction of local features on non-uniformly distributed point cloud data and poor classification and segmentation results caused by ignoring some point cloud features. An inverse density point cloud recognition and segmentation algorithm D-PointNet++ based on PointNet++ and density information is proposed. The point cloud density was used to calculate the sampling probability of each point, and the multinomial distribution was used to sample the point cloud according to the sampling probability. The point cloud was grouped by adaptively scaling the grouping radius. A variety of pooling methods were used to extract point cloud features and the weights of multiple features were calculated by using the multi-head attention mechanism, and the global features of the point cloud were obtained by weighted aggregation. Experimental results show that, compared with a variety of evaluation algorithms, D-PointNet++ has a significant improvement in point cloud classification accuracy and segmentation accuracy.