LiDAR Place Recognition Based on Range Image and Column-Shift-Invariant Attention

LiDAR place recognition is a crucial component for the robust localization of intelligent vehicles with LiDAR sensors. The existing methods based on range image often resort to the permutation invariance of NetVLAD for rotation robustness, which restricts the change of feature scales and affects the recognition performance. In this paper, a LiDAR place recognition method based on range image and column-shift-invariant attention is proposed. Instead of relying on the invariance of NetVLAD, we design an attention mechanism whose attention map shifts synchronously with the input features. Under the weighted form of matrix multiplication, the attention module achieves the invariance to the cyclic column shift of range image while capturing the global contextual information. This enables the variation of feature scales. On this basis, a multi-scale module with feature downsampling and mixing is presented to mine multi-scale information after the attention. It enhances the discriminability of the global descriptor and further facilitates place recognition performance. The effectiveness of proposed method is evaluated on the KITTI, Ford Campus, and NCLT datasets.