【摘要】 由于激光雷达等三维扫描设备分辨率限制、目标间的相互遮挡以及目标表面材质透明等问题，采集到的三维点云数据往往是不完整的．近年来，以数据驱动为主的深度学习方法逐渐被用于解决点云补全问题，然而，现有的点云补全算法致力于补全出原始目标点云的整体拓扑结构而忽略了对于目标点云局部细节位置的恢复．针对这一问题，提出了一种基于特征融合的深度学习点云补全算法，利用关系感知编码器提取得到的点云邻域分布特性和点云空间特征进行融合并把点云映射为128维的潜在特征向量，接着通过树状解码器将全局特征向量与局部特征向量进行特征融合并分形输出补全的点云预测．在开源数据集ShapeNet和KITTI上的仿真结果显示，提出的点云补全算法在补全精度和可视化效果上相比于现有主流的点云补全算法均有了提升．在相互倒角距离指标上，本文算法在飞机、汽车、椅子和桌子点云模型上分别减少了2.4%、2.5%、8.3%和6.2%；在单向豪斯多夫距离指标上，本文算法在飞机、汽车、椅子和桌子点云模型上分别减少了2.1%、0.4%、0.8%和0.3%．最后，利用镭神智能16线激光雷达实测点云数据进行点云补全实验验证了所提出的点云补全算法的有效性，说明了在ShapeNet数据集中训练的点云补全网络能够很好地适用于激光雷达测量得到的稀疏点云数据．更多还原

【Abstract】 Owing to the limitation of 3 D scanning devices such as lidar,mutual occlusion among targets,and the transparency of surface materials,the acquired 3 D point cloud is often incomplete. Recently,data-driven deep learning methods have been gradually applied to complete point clouds. However,existing point cloud completion methods only focus on completing the global topological structure of the original point cloud and ignore the local detail recovery of the target point cloud. To address this limitation,this study proposes a novel deep learning point cloud completion algorithm via feature fusion. This algorithm uses a relation-aware encoder,which performs the mapping from the point cloud to the 128-dimensional latent feature vector to fuse the neighbor distribution characteristics and spatial features of the point cloud. Subsequently,a tree-structured decoder is applied to fuse the features of the global and local feature vectors to output the complete point cloud predictions. The simulation results of the ShapeNet and KITTI datasets show that the proposed point cloud completion algorithm exhibits better performance than state-of-theart methods on completion precision and visualization. The proposed method reduces the chamfer distance on the airplane,car,chair,and table point cloud models by 2.4%,2.5%,8.3%,and 6.2%,respectively. In contrast,the proposed method reduces the unidirectional Hausdorff distance on the airplane,car,chair,and table point cloud models by 2.1%,0.4%,0.8%,and 0.3%,respectively. Finally,the point cloud completion experiment on the actual scanning of point clouds measured using the LeiShen intelligent system lidar verifies the effectiveness of the proposed point cloud completion algorithm,which indicates that the point cloud completion networks trained on the ShapeNet dataset are suitable for sparse point clouds measured using lidar.更多还原