Patchlpr: a multi-level feature fusion transformer network for LiDAR-based place recognition

LiDAR-based place recognition plays a crucial role in autonomous vehicles, enabling the identification of locations in GPS-invalid environments that were previously accessed. Localization in place recognition can be achieved by searching for nearest neighbors in the database. Two common types of place recognition features are local descriptors and global descriptors. Local descriptors typically compactly represent regions or points, while global descriptors provide an overarching view of the data. Despite the significant progress made in recent years by both types of descriptors, any representation inevitably involves information loss. To overcome this limitation, we have developed PatchLPR, a Transformer network employing multi-level feature fusion for robust place recognition. PatchLPR integrates global and local feature information, focusing on meaningful regions on the feature map to generate an environmental representation. We propose a patch feature extraction module based on the Vision Transformer to fully leverage the information and correlations of different features. We evaluated our approach on the KITTI dataset and a self-collected dataset covering over 4.2 km. The experimental results demonstrate that our method effectively utilizes multi-level features to enhance place recognition performance.