VPD-Map: Visual Pointcloud Descriptor with Top-view Feature Map for Viewpoint-Free Loop Closure in Direct Visual Odometry

Abstract

Loop closing is a crucial step to improve the accuracy of a SLAM system, which can reduce drift caused by frontend odometry and ensure global consistency of mapping process. In a traditional visual SLAM system, this is usually done by extracting and matching image features. The accuracy of the loop closure detection and the precision of the pose-constraint provided by loop closing is the bottleneck restricting the loop closing accuracy. Since image features is sensitive to large viewpoint changes, in this paper, we propose a novel neural network architecture VPD-Map, which takes visual pointcloud to provide a global 3D visual descriptor for fast loop closure detection and a top-view feature map for pose-constraint prior. Since the descriptor and feature map are based solely on visual pointcloud information, it is robust to viewpoint changes. VPD-Map can also serve as the backend of a visualSLAM system. Experiment on loop closure detection shows that this descriptor can perform viewpoint-free loop closure detection and outperforms traditional loop detection algorithms like bag-of-word model on the KITTI Odometry Dataset.