Rotation-robust LiDAR-based Place Recognition using Multi-channel Scan Context and Representation Learning

Abstract

I present a study focused on LiDAR-based place recognition, aiming to enhance robustness against rotation through the multi-channel scan context and instance-wise contrastive learning. In order to achieve this, a multi-channel scan context comprising three components was constructed: the geometric channel for representing geometric structures, the semantic channel for encoding semantic information, and the intensity channel for capturing intensity data. This multi-channel representation enables more effective spatial characterization compared to single-channel descriptors commonly used in existing methods.

Furthermore, instance-wise contrastive learning, a form of representation learning, was employed to train a feature extraction neural network. Unlike conventional pair-wise learning approaches, our method reveals the overall relationship between descriptors in the representation space. This effectively captures the global relationship and shape similarity between descriptors. Experimental evaluations conducted on the KITTI dataset demonstrate the superior performance of our proposed approach in terms of place recognition accuracy, as well as its robustness to rotation and translation.

In summary, this research improve the robustness of LiDAR-based place recognition by combining a multi-channel scan context with instance-wise contrastive learning. This work holds great potential for various applications in autonomous driving, specially mapping and localization.