Robust point cloud registration utilizing learning features and graph attention network

Abstract

Point cloud registration is a central function in many utilizations such as localization, mapping, tracking, and reconstruction. Successful registration relies on the extraction of powerful and discriminatory geometric features. Although existing learning-based methods require high computational power to process a large number of raw points at the same time, the computational power limitation is not an issue thanks to parallel GPU usage. This thesis introduces an efficient and economical extraction framework of Dense Features using a Graph Attention Network for point cloud registration and matching. The detector is responsible for finding high-confidence keypoints in large raw data sets. The descriptor takes these key- points combined with their neighbors to extract density-invariant features to prepare for the match. The Graph Attention Network uses an attention mechanism that enriches the relationships between point clouds. Finally, we treat this as an optimal transport problem and use the Sinkhorn algorithm to find positive and negative matches. We perform thorough tests on the KITTI dataset and evaluate the effectiveness of this approach. The results show that this method with efficient compact keypoint selection and delineation can achieve the best performance matching indices and achieve the highest success rate of 99.88% registration compared to with other modern approaches.