LiDAR-based lifelong robotic mapping in changing environments

Abstract

This thesis explores lifelong robotic mapping using light detection and ranging (LiDAR) sensor. Lifelong mapping means long-term map management after the map creation of a single moment. A lifelong mapper this thesis proposes is a modular system that integrates an overall pipeline from multiple map creation, long-term place recognition, aligning multiple maps, change detection, to live/meta map maintenance. Such integration of each advanced module is particularly critical for autonomous vehicles in outdoor environments, where 3D changes cause a pre-built map to be out-dated are inevitable. The first part of the thesis proposes a novel robust LiDAR-based place recognition method, named Scan Context++. Place recognition is the most core component not only for accurate robot map building to resolve robot motion drifts but also for aligning multiple disjoint sessions in a shared frame. Scan Context++ possesses multiple invariances robust to external and internal changes. The invariances make Scan Context++ effectively performs even under long-term temporal gaps (e.g., 1 year). Scan Context++ outperforms existing state-of-the-art LiDAR place recognition methods. Also, it runs in real-time on a CPU, nearly 100 Hz for open large-scale urban datasets. The experiments are extensively conducted for multiple sequences of multiple cities that include various loop candidate situations. In the second part of the thesis, an open and modular framework for LiDAR-based lifelong mapping (LT-mapper) is proposed. To achieve lifelong map management, three subproblems sequentially to be conquered. First is accurate single-session map creation and aligning multiple single-session maps. To do so, Scan Context++, proposed in the first part, is integrated with state-of-the-art LiDAR odometry to build a complete LiDAR SLAM system. Thus, we construct the full-SLAM system and release it as an open source, named SC LiDAR SLAM and LT-SLAM for multi-session SLAM. Second, change detection between two aligned maps should be made. Change is divided into two categories, high dynamic, and low dynamic. A range image-based removing-and-reverting algorithm is proposed to deal with both types of changes. The corresponding modules will be introduced as Removert and LT-removert. The method first introduces reverting mechanism and successfully reducess the number of falsely removed static points. In addition, LT-removert supports a function that can automatically segregate dynamic object points without human supervision by matching raw scans from the made static maps. The last subproblem is to integrate the aforementioned modules seamlessly. The integrated framework is LT-mapper with a set of readily available APIs as an open-source. Finally, an efficient (in both memoery and time cost) change management methods named LT-map is proposed. As the last part of the thesis, the thesis is finished by analyzing the contributions and limitations of Scan Context++ and LT-mapper.