This thesis is concerned with the problem of creating three-dimensional terrain maps using a high-density, high rate LiDAR sensor mounted on a moving platform without the use of an external localization solution. In particular, the thesis is motivated by the problem of building terrain maps in open-pit mining environments from data produced by haul-truck mounted LiDAR sensors.
A foundation problem in map building from such data is the process of assembling together the point clouds generated by individual sensor scans into a common frame of reference. This falls into a broader class of well-studied problems called scan matching. Scan matching is most commonly solved by the Iterative Closet Point (ICP) method.
This thesis makes two contributions. The first contribution is a comparison of the different ICP variants for the candidate application. The literature is replete with different algorithms that can be used at different stages of the ICP process. A population of 20,736 ICP-variants drawn from methods proposed in the literature is compared. Whilst others have looked to compare different ICP-variants, this investigation is distinguished from these earlier works in three respects: i) its comprehensiveness; ii) the focus of its target application (the loosely structured terrain of open-cut mining); and iii) its emphasis on how accuracy, precision, and computational efficiency trade-off across different variants. The main finding of this investigation is that the geometry of the point cloud critically determines the quality of the scan match. Significantly, of the variants considered, none were found to simultaneous meet requirements on accuracy, precision, and efficiency, highlighting the need for better approaches.
The second contribution comes in the form of progress towards improving on the performance of established methods. For this, the thesis introduces the concept of “eigentropy” to quantify the geometric disorder or geometric information for points of a cloud. Eigentropy is conceptually similar to entropy as it appears in thermodynamics and information theory. Three novel methods are introduced to improve the performance of ICP-based scan-matching for loosely structured terrain. These methods are termed the: eigentropy filter, matching by normal deviation and unilateral eigentropy rejection, and they are based on the geometrical information content at a point. The results associated with these methods improve the overall performance of the ICP algorithm and five accurate, precise, efficient, and robust ICP variants are identified from an expanded population of 73,728. These five variants all use the methods introduced in this thesis.