Abstract: To address the problem that the accuracy of pose estimation was low due to the existence of numerous similar features and non-corresponding points in the point cloud, a global feature pose estimation method based on keypoint distances was proposed. This method constructed global features using the distances between keypoints, avoiding the influence of similar local features on the accuracy of pose estimation. Meanwhile, to improve the matching speed of global features, a feature-matching strategy based on a distance comparison table was proposed. Similarity measurement was carried out by voting on global features through the comparison table. Thus, while avoiding the interference of non-corresponding points, the efficiency of finding the correspondence of global features was effectively enhanced. Finally, the identified correspondences were subjected to a new outlier removal strategy based on reliability of correspondence graph (GROR) to eliminate outliers, and the transformation pose was obtained. Experiments were conducted on four public datasets. The results show that, compared with the Fast Point Feature Histogram (FPFH), the Signature of Histograms of Orientations (SHOT), and the Binarized Signature of Histograms of Orientations (BSHOT), the area under the precision-recall curve of the feature matching of the proposed method is increased by 116%, 169%, and 137% respectively. Moreover, compared with the original GROR, the rotational error and translational error of the proposed method are reduced by 47.38% and 52.43% respectively.

Key words: machine vision, six degrees of freedom pose estimation, keypoint distance, global feature, distance comparison table

摘要： 为解决位姿估计中由于点云存在较多的相似特征和非对应点导致位姿估计精度低的问题，提出一种基于关键点距离的全局特征位姿估计方法，该方法使用关键点之间的距离构建全局特征，避免相似的局部特征对位姿估计精度的影响。同时，为提升全局特征匹配速度，提出了一种基于距离对照表的特征匹配策略，通过对照表对全局特征投票进行相似度量，从而在避免非对应点干扰的同时，有效地提高了通过全局特征找寻对应关系的效率。最后，将这些对应关系使用基于对应图可靠性的外点去除策略(GROR)去除外点并得到转换位姿。通过在四个公开数据集中进行实验，结果显示，相较于快速点特征直方图(FPFH)、方向直方图签名(SHOT)和二值化方向直方图签名(BSHOT)三个特征描述子，所提方法在特征匹配的精度-召回率曲线的下区域面积指标分别提升116%、169%和137%。以及相较原GROR，所提方法在旋转误差和平移误差上分别降低47.38%和52.43%。

关键词: 机器视觉, 六自由度位姿估计, 关键点距离, 全局特征, 距离对照表