Point cloud based loop detection and localization

Abstract

Localization is one of the most essential elements for autonomous vehicles because autonomous navigation totally relies on the awareness of the current location. SLAM is an important technique for lots of localization methods, while most of the SLAM methods would suffer from drift in long-time processing. Loop closure detection is an effective method to correct the long-time drift and improve the result of SLAM. In this project, leveraging on the powerful capability of PointNetVLAD, a method used for loop closure detection was proposed, and it can serve as an important technique in dealing with the drift problem in SLAM. As well, two retrieval mechanisms, temporal consistency retrieve (TCR) and temporal spatial consistency retrieve (TSCR), were proposed to improve the localization performance based on the original PointNetVLAD method. The results show that the loop detection method is feasible and has good performance. As well as, the introduction of TCR and TSCR mechanism can largely improve the accuracy of localization, compared with that using primitive PointNetVLAD recall at 1.