AttDLNet: Attention-Based Deep Network for 3D LiDAR Place Recognition

Place recognition has often been incorporated in SLAM and localization systems to support autonomous navigation of robots intelligent vehicles. With the increasing capacity DL approaches learning useful information from 3D LiDARs, place also benefited this modality, which led higher re-localization loop-closure detection performance, particularly, environments with significantly changing conditions. Despite progress field, efficient extraction invariant descriptors LiDAR data is still a challenging problem domain. In work, we propose novel LiDAR-based deep network that resorts self-attention mechanism to, on one hand, leverage computational efficiency these operations and, other, reweigh relevant local features thus create discriminative descriptors. The proposed trained validated KITTI dataset an ablation study presented assess components network. Results show adding attention improves leading loop closures, outperforming established approach. From study, results indicate middle encoder layers have highest mean while deeper are more robust orientation change. code publicly available at: https://github.com/Cybonic/AttDLNet