Towards Stable Adversarial Feature Learning for LiDAR based Loop Closure Detection

TL;DR

This paper introduces a novel stable adversarial feature learning method using GANs for LiDAR-based loop closure detection in SLAM, addressing training instability and improving robustness against viewpoint changes.

Contribution

It is the first to extract adversarial features from LiDAR data, combining an inner cycle restriction and side updating modules to enhance training stability and robustness.

Findings

Features are more stable during training.

Significantly improves robustness to viewpoint variations.

Outperforms state-of-the-art methods on KITTI dataset.

Abstract

Stable feature extraction is the key for the Loop closure detection (LCD) task in the simultaneously localization and mapping (SLAM) framework. In our paper, the feature extraction is operated by using a generative adversarial networks (GANs) based unsupervised learning. GANs are powerful generative models, however, GANs based adversarial learning suffers from training instability. We find that the data-code joint distribution in the adversarial learning is a more complex manifold than in the original GANs. And the loss function that drive the attractive force between synthesis and target distributions is unable for efficient latent code learning for LCD task. To relieve this problem, we combines the original adversarial learning with an inner cycle restriction module and a side updating module. To our best knowledge, we are the first to extract the adversarial features from the light detection and ranging (LiDAR) based inputs, which is invariant to the changes caused by illumination and appearance as in the visual inputs. We use the KITTI odometry datasets to investigate the performance of our method. The extensive experiments results shows that, with the same LiDAR projection maps, the proposed features are more stable in training, and could significantly improve the robustness on viewpoints differences than other state-of-art methods.