I. Introduction
Simultaneous localization and mapping (SLAM) [1] plays a significant role in the field of autonomous driving [2]. Since the global navigation satellite system can easily be affected by environmental occlusion and multipath effects while the map-based localization stays robust [3], most autonomous vehicles depend on the localization information provided by SLAM [4]. Although drifts are inevitable in the entire SLAM processing and usually militate against the estimated state accompanied by trajectory [5], the impact of drifts can be effectively eliminated by visiting the same place multiple times [6] and thereby creating a consistent map. Place recognition is also known as loop detection in SLAM [7], which is critical for identifying and returning to the same position. In the past 20 years, many studies have established vision-based methods and shown the practicability in actual cases. However, the performances of these methods are relatively unstable because of the changes in recurrent viewpoint and transformations in light intensity. In contrast, the active sensor lidar is less susceptible to being generalized for place-recognizing tasks recently.