目的 点云目标识别流程分为离线与在线阶段。离线阶段基于待识别目标的CAD模型构建一个模型库，在线基于近邻查找完成识别。本文针对离线阶段，提出一种新的模型库构建方法。方法 首先将CAD模型置于一个二十面体中心，使用多个虚拟相机获取CAD模型在不同视角下的点云；然后将每个不同视角下的点云进行主成分分析并基于主成分分析的结果从多个选定的方向将点云切分为多个子部分，这些子部分包含点云的全局及局部信息；接着对每个子部分使用聚类算法获取其最大聚类，去除离群点；最后结合多种方式删减一些冗余聚类，减小模型库规模。结果 在多个公开数据集上使用多种点云描述子进行对比实验，识别结果表明，相对于传统的模型库构建方法，基于本文方法进行识别正确率更高，在某些点云描述子上的识别正确率提升达到10%以上。结论 通过将CAD模型在不同视角下点云的全局与局部信息都加入模型库中，本文提出的模型库构建方法可有效提高点云目标识别正确率，改善了场景目标发生遮挡时，近邻查找识别精度不高的问题。

Objective Frameworks for point cloud object recognition are generally composed by two stages. An offline stage constructs a model library, and an online stage recognizes objects by using nearest neighbor search. Traditional methods use global surfaces to construct a model library, which is sensitive to occlusion and inaccurate segmentation result. This study investigates the offline stage and presents a novel model library construction method. Method The proposed method simulates possible occlusions and adds point clouds with simulated occlusions to the model library to alleviate the influence of occlusion and inaccurate segmentation result. First, a CAD model is placed at the center of an icosahedron, and multiple virtual cameras are used to obtain the partial point clouds of the model. For each partial point cloud, a local coordinate system is constructed using principal component analysis, and the point cloud is aligned with the coordinate system. This process makes the proposed method invariant to rigid transformations. Second, several direction vectors are obtained based on the local coordinate system, and the partial point clouds are segmented into multiple subparts based on the length of the point cloud on each direction vector. Simulation of occlusion at different degrees is performed on these subparts, which contain the global and local surfaces of the partial point cloud. Third, a simple clustering method is used to obtain the largest cluster of the subparts, and outliner points are removed at this stage. The largest cluster will be added to the model library only if the cluster has sufficient points. This process reduces the memory requirements and decreases time consumption during the nearest searches. Redundant clusters with similar surface in the library are still observed after removing the clusters with few points. Finally, an iterative closest point(ICP) based algorithm is used to remove the point clouds with similar surfaces, thereby further decreasing the memory requirements. Subsequently, only dozens of subparts are used to describe each of the CAD model. Result Experimental results on two public datasets show that the proposed method promotes recognition accuracy at different levels. For the UWAOR dataset, the recognition performance on five types of point cloud descriptor is remarkably improved. Particularly, the proposed method enhances the recognition performance by 0.208 on the GASD descriptor and 0.173 on the ROPS descriptor (k=1 in KNN). For the Bologna Random Views dataset, the proposed method enhances the recognition accuracy of most of the point cloud descriptors. For example, the proposed method improves the recognition rate by 0.193 (k=1) for the GASD descriptor. However, the recognition improvement on Bologna Random Views dataset is slightly lower than that of the UWAOR dataset. This condition is partially caused by the lighter occlusion of scene objects on Bologna Random Views dataset compared with the UWAOR dataset. Experiments at different noise levels are also conducted. The noise that follows a Gaussian distribution with different variances and zero means are added to the scene point cloud. Experimental results show that the proposed method maintains the recognition rate promotion with the increase on the standard deviation of noise. For example, the proposed method enhances the recognition rate of ESF descriptor by 0.162 (no noise) and 0.034 (noise with a standard deviation of 3×mesh resolution) for the UWAOR dataset. This finding can be interpreted as the subparts having considerable points to overcome the influence of noise. Conclusion The proposed method enhances the recognition performance by combining the global and local surfaces of partial point clouds in constructing the model library, especially when the scene objects are occluded or have inaccurate segmentation. This outcome is valuable because it reduces the time consumption of the subsequent hypothesis verification stage. A better redundancy reduction algorithm should be proposed in future studies, in which each of the CAD model can be represented with the same number of subparts. In the present work, various subparts are used to describe different CAD models, which has affected the recognition results of nearest neighbor search. Meanwhile, the coarse pose of the scene object can be estimated by aligning the scene object with the point cloud in the model library, and the ICP-based algorithm can be used to refine the coarse pose to obtain precise pose information.