Multi-channel Scan Context for LiDAR-based Place Recognition Using Siamese Neural Network

Abstract

LiDAR-based place recognition is a key component of LiDAR-based localization. However, due to the unordered, unstructured, and noisy characteristics of point cloud data from LiDAR sensors, there are limitations in using raw LiDAR data for place recognition. Therefore, it is essential to perform conversion processing to generate LiDAR descriptors for place recognition. In this paper, we propose a new method for generating descriptors adopting the feature extraction method of scan context.We attempt to improve the performance of place recognition by implementing a multi-channel scan context that combines geometric, semantic and intensity information. Furthermore, utilizing the rotation-invariant Siamese neural network, we propose a robust descriptor to handle translation, roll and pitch motions. We verify the performance of our descriptor generation and place recognition performance through an experiment using the semantic KITTI dataset.