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

Abstract

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