基于深度学习的高精度点云补全算法

High Precision Point Cloud Completion Algorithm Based on Deep Learning

由于获取三维点云数据设备分辨率低及物体遮挡等原因,不可避免地会导致获取的点云数据通常是不完整的,这对点云分割、点云重建等下游任务会产生巨大的影响。近年来基于深度学习的点云补全算法开始广泛应用并取得较好的效果,但大多数算法只关注点云的全局信息,并没有充分考虑到点云的局部特征,难以表征点云空间复杂的变化关系,所以补全精度上还存在欠缺。该文提出一种基于深度学习的高精度点云补全算法,算法整体采用编码器-解码器结构,创造性地在特征提取模块引入卷积层DOConv,并在特征融合模块添加结合空间注意力机制和通道注意力机制的双重注意力机制,可以融合不同层次的特征。该算法可以灵活地对点云全局和局部特征进行提取并综合关键点的局部关联性和全局信息。实验结果表明,与几个主流算法相比,该算法补全精度更高,可以得到更为完整准确的点云模型。

Due to the low resolution of the device for acquiring 3D point cloud data and the occlusion of objects,the acquired point cloud data is usually incomplete,which will have a huge impact on downstream tasks such as point cloud segmentation and point cloud reconstruction.In recent years,point cloud completion algorithms based on deep learning have been widely used and achieved good results,but most algorithms only focus on the global information of point clouds,and do not fully consider the local features of point clouds,making it difficult to characterize the point cloud space.There is a complex change relationship,so there is still a lack of completion accuracy.We propose a high-precision point cloud completion algorithm based on deep learning.The algorithm adopts the encoder-decoder structure as a whole,creatively introduces the convolutional layer DOConv in the feature extraction module,and adds combined spatial attention in the feature fusion module.The dual attention mechanism of force mechanism and channel attention mechanism can fuse features at different levels.The algorithm can flexibly extract global and local features of point clouds and synthesize local correlation and global information of key points.The experimental results show that compared with several mainstream algorithms,the proposed algorithm has higher completion accuracy and can obtain a more complete and accurate point cloud model.