Large-scale multi-session point-cloud map merging

Abstract

This thesis introduces LAMM, an open-source framework for large-scale multi-session 3D LiDAR point cloud map merging. LAMM can automatically integrate sub-maps from multiple agents carrying LiDARs with different scanning patterns, facilitating place feature extraction, data association, and global optimization in various environments. Our framework incorporates two key novelties that enable robust, accurate, large-scale map merging. The first novelty is a temporal bidirectional filtering mechanism that removes dynamic objects from 3D LiDAR point cloud data. This eliminates the effect of dynamic objects on the 3D map model, providing higher-quality map merging results. The second novelty is a robust and efficient outlier removal algorithm for detected loop closures. This algorithm ensures a high recall rate and a low false alarm rate in position retrieval, significantly reducing outliers in repetitive environments during large-scale merging. We evaluate our framework using various datasets, including KITTI, HeLiPR, WildPlaces, and a self-collected colored point cloud dataset. The results demonstrate that our proposed framework can accurately merge maps captured by different types of LiDARs and data acquisition devices across diverse scenarios.