遮挡条件下基于MSF-PPD网络的绿萝叶片点云补全方法

针对在自然场景中,由于遮挡、视角限制和操作不当等问题,导致传感器获取的植物或器官点云不完整,提出了一种基于多尺度特征提取模块结合点云金字塔解码器(Multi-scale feature extraction model with point cloud pyramid decoder,MSF-PPD)的叶片形状补全网络。首先,采用多尺度特征提取模块实现不同维度特征信息的全局提取和融合,其次,通过点云金字塔解码器进行叶片点云的多阶段生成补全,最终得到完整的目标叶片形状。使用曲面参数方程构建绿萝叶片仿真模型库,并将其离散成点云作为网络模型训练的训练集和验证集,使用Kinect v2相机获取绿萝叶片点云作为网络模型补全性能评估的测试集。试验结果表明,本文网络结构对叶片点云补全的效果理想,证明本文方法能够对遮挡情况下的绿萝叶片进行高效、完整的补全。

Point Cloud Complementation Method of Epipremnum aureum Leaves under Occlusion Conditions Based on MSF-PPD Network

For the natural scenes, the point clouds of plants or organs acquired by sensors are incomplete due to the problems of occlusion, viewpoint limitation and improper operation. A multi-scale feature extraction model with point cloud pyramid decoder (MSF-PPD) network was proposed for leaf shape complementation. Firstly, the multi-scale feature extraction module was used to achieve the global extraction and fusion of different dimensional feature information, and secondly, the multi-stage generation of leaf point cloud was complemented by the point cloud pyramid decoder to finally obtain the complete target leaf shape. A library of Epipremnum aureum leaf simulation models was constructed by using surface parametric equations and discretized into point clouds as the training set and validation set for network model training, and the Epipremnum aureum leaf point clouds were obtained by using the Kinect v2 camera as the test set for model complementary performance evaluation. The experimental results showed that the network structure had an ideal effect on leaf point cloud complementation, which proved that the method proposed was able to perform efficient and complete complementation of Epipremnum aureum leaf under the obscured situation.