PointAttentionVLAD: A Two-Stage Self-Attention Network for Point Cloud-Based Place Recognition Task

Abstract

Point cloud-based place recognition plays a crucial role in robotics and unmanned vehicle tasks, particularly in relocalization and loop detection modules of LiDAR-based simultaneous localization and mapping systems. It’s also essential for global localization in scenarios where prior pose information is unavailable. However, three-dimensional point cloud data are characterized by sparsity and disorder, making it challenging to extract robust features. This study proposes an end-to-end deep learning network to compress the point cloud into a global descriptor for point cloud retrieval tasks. The proposed network implements two self-attention modules, i.e., the local point cloud-based self-attention and global point cloud-based self-attention mechanisms. Due to this two-stage self-attention mechanism, the proposed PointAttentionVLAD network achieved a higher average recall @ top 1 on the Benchmark datasets than the SOE-Net and LPD-Net algorithms by 0.39% and 3.41%, respectively. Furthermore, experiments were conducted on KAIST dataset to assess the generalization ability of the proposed PointAttentionVLAD, and the proposed network demonstrated impressive performance on KAIST dataset. The code and the pre-trained model of the proposed PointAttentionVLAD are available at https://github.com/leo6862/pointattentionvlad.

Keywords

• Simultaneous localization and mapping (SLAM)
• point cloud place recognition
• loop detection
• deep learning
• relocalization