AttDLNet: Attention-based DL Network for 3D LiDAR Place Recognition

TL;DR

AttDLNet introduces an attention-based deep learning model for 3D LiDAR place recognition, improving robustness to environmental changes and orientation variations, and outperforming existing methods in loop closure detection.

Contribution

This work presents a novel attention-based deep learning architecture for 3D LiDAR place recognition, utilizing range-based proxy representations and stacked attention layers, with demonstrated superior performance.

Findings

Attention improves place recognition performance.

Middle encoder layers yield highest accuracy.

Deeper layers are more robust to orientation changes.

Abstract

LiDAR-based place recognition is one of the key components of SLAM and global localization in autonomous vehicles and robotics applications. With the success of DL approaches in 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 significant changing conditions. Despite the progress in this field, the extraction of proper and efficient descriptors from 3D LiDAR data that are invariant to changing conditions and orientation is still an unsolved challenge. To address this problem, this work proposes a novel 3D LiDAR-based deep learning network (named AttDLNet) that uses a range-based proxy representation for point clouds and an attention network with stacked attention layers to selectively focus on long-range context and inter-feature relationships. The proposed network is trained and validated on the KITTI dataset and an ablation study is presented to assess the novel attention 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. The code is publicly available at https://github.com/Cybonic/AttDLNet