HP-KPConv: Point Cloud Registration Evaluation via Local Feature Distribution Similarity

ICP is one of the most popular solutions for point cloud registration. However, the evaluation metric of the ICP algorithm, based on point-wise Euclidean distances, falls short in meeting the demands of high-precision alignment. In this paper, we propose a point cloud registration evaluation method based on local feature distribution to assess alignment quality, in particular the translation errors among matched point clouds. Rather than comparing the point-to-point Euclidean distance as adopted in ICP, our main idea is that the spatial distribution of one local region pre- and post-alignment should be similar. We design a filter to calculate local region similarity and find out outliers according to the similarity. This high-precision filter is integrated into the KPConv network, denoted as HP-KPConv. Finally, after conducting extensive experiments on Dingo robots and LiDAR, we validate the ability of HP-KPConv to effectively detect subtle alignment errors, proving its applicability in robot localization evaluations.