Sustainable Globally Aligned Lidar Point Cloud Mapping 光達點雲地圖的規模化建置

Introduction
Lidar point clouds let us create precise 3-d maps of roads, signs, utility poles, powerlines, buildings, etc., enabling new possibilities in applications such as autonomous vehicles, urban planning, augmented reality, land administration, and infrastructure asset maintenance.
A Lidar point cloud is created by driving a scanner-equipped vehicle through an area and using the scanner's trajectory to transform Lidar distance and angle measurements into 3-d points.  But errors in the trajectory can cause misalignments between the point clouds created by separate drives through the same area.  Even trajectories measured using high end GNSS+INS can easily have absolute errors in the tens of centimeters, particularly in urban environments with limited sky visibility and multipath effects; this level of error is too large for aligning thin objects such as utility poles.
Future 3-d mapping efforts will involve frequent collecting of large numbers of overlapping drives.  We need a sustainable way of maintaining global alignment of this continually growing set of drives.
Many surveying and mapping companies are already offering services for creating high quality Lidar point clouds of a given project area.  Alignment within a project is usually very precise, thanks to the use of high end GNSS+INS equipment, manually selected tie points, and automatic plane matching solutions such as Riegl RiPROCESS.  But alignment across projects is usually less precise, because it is only indirectly enforced via GCPs.  The large number of projects that get created over time present huge opportunities for change detection, seamless 3-d modelling over multiple project areas, and filling in of missing parts e.g. due to parked cars, if only we have a sustainable way of maintaining global alignment across all projects.
Increasingly, tech companies such as DeepMaps, MobilEye and Atlatec are leveraging low cost sensors deployed at a large scale to create maps with greater coverage and freshness. This can nicely complement the accuracy of survey-quality maps, but will require even greater levels of automation to align in a cost-effective fashion.
Automatic creation of aligned point clouds has been studied by many papers in robotics and SLAM (see survey paper by Cadena et.al. [1]).  These methods use scan matching to estimate relative poses, which are then combined with IMU and GNSS measurements in a pose graph optimization [2] to obtain a trajectory, a trajectory that can be used to create a point cloud that is aligned with quality commensurate with the quality of scan matching.  Existing automatic methods include Cartographer [3] and LIO-SAM [4], which focus on the single drive case; the paper by Ding et.al. [5], which considers alignment of multiple drives; and perhaps the largest scale demonstration in a paper to date by Shiratori et.al. [6], who automatically created an aligned point cloud for the entire city of San Francisco.
However, sustained automation at a large scale still has many challenges.  This includes long-term maintenance of pose graphs [7], optimizing large pose graphs [8], and robustness to scan matching mistakes [9].  While these papers focus on full autonomy, we are willing to do manual edits, as long as the overall approach is sustainable.
Sustainability is not only about how large of a geographic area, but also how frequently a map can be updated, how many person-hours for manual edits, and how much compute power and time for automated processing.  For example, can we achieve sustained weekly mapping of the entire Hsinchu city (~100 km2) using a single full-time worker and a single cluster of computers?  In this article, we describe our initial efforts (Figure 1) towards building such a system.

Figure 1  Coverage of point cloud map created so far by our system.