| 摘要: |
| 针对现有的稠密点云配准方法依赖初始位置设定、计算成本高、配准成功率不高等问题，提出了一种基于点云局部几何特征的稠密点云配准方法。采用深度卷积网络模型提取点云的局部几何特征，从而减少了三维点云数据的噪声、低分辨率和不完备性等带来的影响。在此基础上，使用K维树搜索完成局部几何特征描述子的关联工作。最后，通过随机采样一致算法对点云的相对位姿进行鲁棒的估计。通过对开源数据集上5个典型场景中的数据测试表明，该方法的配准成功率达到92.5%，配准精度达到0.0434m，配准时间相对最邻点迭代配准算法缩短了74.7%，实验结果验证了该方法的有效性、实时性和鲁棒性。 |
| 关键词: 室内定位 三维点云配准 三维点云描述子 深度学习 |
| DOI： |
|
| 基金项目:国家自然科学基金（61873163）;装备预研领域基金（61405180205，61405180104） |
|
| A Dense Point Cloud Registration Method Based on Local Geometry Features |
| ZHU Yi-fan,PEI Ling,WU Qi,XIA-Song Pengcheng,LI Tao,CHEN Lei,YU Wen-xian |
| (Shanghai Key Laboratory of Navigation and Location-based Services, Shanghai Jiao Tong University, Shanghai 200240, China;Shanghai Key Laboratory of Navigation and Location-based Services, Shanghai Jiao Tong University, Shanghai 200240, China; Beijing Institute of Tracking and Telecommunication Technology, Beijing 100094, China) |
| Abstract: |
| In order to solve the problems of existing dense point cloud registration methods, such as the dependence on initial location setting, high computational cost, and low registration success rate, a dense point cloud registration method based on local geometric features of the point cloud is proposed. A deep convolutional neural network is used to extract local features of a point cloud, to reduce the influence of noise, low resolution, and incompleteness of 3D point cloud data. On this basis, a k-dimensional tree search is used to complete the association of local geometric feature descriptors. Finally, the random sampling consensus algorithm is used to estimate the relative pose of the point cloud robustly. The experimental results by testing 5 typical scenarios from open-source data sets show that the registration success rate and the accuracy of the proposed method reach 92.5%, 0.0434m, and the registration time is 74.7% shorter than the nearest point iterative registration algorithm. Experimental results show that the proposed method is effective, real-time and robust. |
| Key words: Indoor localization 3D point cloud registration 3D point cloud descriptor Deep learning |
