A Novel Place Recognition Network using Visual Sequences and LiDAR Point Clouds for Autonomous Vehicles

Place recognition plays an important role in autonomous vehicles localization, particularly in GNSS-degraded environments. LiDAR-based place recognition (LPR) could achieve localization by comparing on-line LiDAR point clouds with a pre-built off-line point-cloud database. However, LiDAR sensors are expensive, which hinders their large-scale deployment on every vehicle. To alleviate this issue, we propose a novel cross-modal network, which replaces on-line point clouds with on-line images captured by a low-cost and lightweight monocular camera. We use image sequences instead of single images, which would be helpful to eliminate false matches since image sequences capture more environmental information. Furthermore, we propose an image sequence descriptor to represent the observed environment by learning multi-image integration and global representation. Experiments on 6 trajectories of the KITTI dataset demonstrate our effectiveness and superiority over single image-based methods.