基于PointNet的车身分割方法

Segmentation method of car body based on PointNet

多视图参数化重构方法对车身拓扑优化结果进行重构时需将车身分割为多个独立部分,以达到分割后拓扑优化车身各部分可独立进行重构的目的。对以车身点云作为输入的分割方法进行研究,采用深度学习理论对PointNet网络框架进行改进,根据车身拓扑优化结果,基于多视图参数化重构方法将车身分割为多个部分并赋予标签,并使用CAD(computer aided design)制作及点云增强操作得到车身分割点云数据集,通过梯度下降算法对网络模型进行训练。最终,车身分割网络模型在车身点云测试集中的准确率为87.7%。

The multi-view parametric reconstruction method needs to divide the car body into multiple independent parts to reconstruct the result of car body topology optimization,so that each part of the topology optimized car body can be reconstructed independently after segmentation.The intelligent segmentation method of car body point cloud as input is studied,and the deep learning theory is used to improve the PointNet network framework.The car body is divided into several parts and given labels based on multi-view parametric reconstruction method according to the body topology optimization result contents.The car body segmentation point cloud data set is obtained by CAD(computer aided design)production and point cloud enhancement operation,and the network model is trained by gradient descent algorithm.Finally,the car body segmentation network model achieves 87.7%accuracy in the car body point cloud test set.