Solving the point cloud registration based on deep learning

Abstract

Point cloud registration is used to create a complete environment map with multiple pieces of scans. It is used mostly to guide autonomous robots when the surrounding is complicated. However, the problem is hard to solve so deep learning networks are often utilized. Most neural networks proposed need manual annotation and selection of patches which is time-consuming and makes the model only work for one particular scenario. In this work, a new method is studied in which a self-supervised learning technique is incorporated into the network. The input to this network is one raw point cloud without any labels or annotations, and as a result, the need for annotating or selecting patches is removed. A key point sampling process is also implemented in this network, which filters non-relevant points and further improves the performance. Based on our experiments, this self-supervised model has the capability of performing better than those requiring manual processing of the input.