LprViT:An Efficient ViT for LiDAR-Based Place Recognition Using Range Image

Place recognition is a crucial component in the realm of simultaneous localization and mapping (SLAM) and autonomous driving. It can also enhance the performance of loop closure detection and global localization. LiDAR-based place recognition usually relies on global feature descriptors to accurately express the similarity between LiDAR scans. However, it is time-consuming to extract and generate the global feature descriptors from the original LiDAR data. As a result, treating LiDAR-based point clouds place recognition as a 2D image problem using range projection is becoming an increasingly common and successful strategy. Vision Transformer (ViT) is a Transformer model based on self-attention mechanism for image classification. Through self-attention mechanism, ViT out-performs CNN at modeling the global information of images. We investigate that whether ViT can be transferred to LiDAR-based place recognition tasks. Firstly, we introduce a new convolutional stem to replace the original linear embedding layer in ViT. Secondly, we establish synergy by integrating low-level fine-grained features from new convolution layer with high-level coarse-grained features from ViT by skip connections. Thirdly, we aggregate these features to construct the global feature descriptor via the NetVLAD technique. The results of experiments on KITTI confirm that our method can operate online and achieve superior performance on par with existing approaches.