preventivo y correctivo Mayor "A"
VERSIÓN DEL
A point cloud representing a section of the Maltese archipelago acquired using airborne LiDaR scanners with around 4 samples per m2 is used in this section.
The data was made available by the institute for climate change and sustainable development at the University of Malta. Figure 5.44 illustrates segmentation results, with the top left representing the raw position only data and the top left representing all generated segments. The bottom row illustrates the edge·*
and surface·* segments on the left and right hand side respectively. The point cloud contains a variety of geographic features, including urban areas, trees and fields. Edge segments are more concentrated in the urban area, whereas the ma- jority of large surface segments are located in the fields to the north of the point cloud. The rubble walls which are used to divide these fields are labelled as edge points and effectively contribute towards delineating a considerable number of fields. Figure 5.44 shows a section of the point cloud showing the segmentation of agricultural fields. Edge points in the urban section of the point cloud are useful to distinguish between the different non-regular housing units. Note that in the majority of cases, only a few points (around 30) are sampled from each house roof as shown in Figure 5.40 second row. The first row of the same fig- ure illustrates a surface·planar segment representing a football ground, and two
edge segments representing the goal posts (circled). The small segments on the third row, represent trees which are all connected to the same surface·complex
(rendered in green in Figure 5.44, bottom right). Another important surf ace
segment is the one representing the road network, which extends from the urban to the agricultural sections and is shown in Figure 5.41.
bin, one surface
segment
fitted into six surface.planar segments segments for: monitor screen speakers desktop computer keyboard webcam
Figure 5.39: PaRSe applied on first office scene, extracts bin on the ground, heating elements from the walls and a variety of desktop items including webcam, monitors, keyboard and speakers from a desk. RanSaC without region growing (bottom row) produces segments spanning parts of multiple objects.
Figure 5.40: Top row illustrates from left to right, football ground structure raw points, the segment representing the pitch, and the boundary of the pitch. Note how some sam- ples are acquired from the goal posts (circled) which are visible as two edge segments. The middle row illustrates a close-up view of the urban area, showing the extracted individual housing structures. On average each house is represented by around 100 points. The third row, illustrates the edge segments representing trees in the bottom part of the point cloud shown in Figure 5.44
Figure 5.41: The biggest surface segment represent the main road network in the point cloud. Note how this segment spans different terrain elevations. A simple RanSaC plane fitting approach would not have been able to discriminate between points on the road and others within house complexes.
Figure 5.42: Small presence of rubble walls between fields are sufficient for the seg- mentation algorithm to distinguish between fields which are rendered using different colours.
oin t Cloud Structure Graphs 140
Lamp and wiring 1 edge segment
Deer model on shelve 3 segments 3 desk tops,
two planar segments (over-segmentation)
Figure 5.43: PaRSe applied on second office scene, extracts desk lamp and wiring from a desktop, and multiple books and files from a shelving unit. Over-segmentation occurs when producing segments for desktops due to points falling between the desks (top left). Note that two additional surface segments are produced on the desk near the lamp wiring. A deer model located on one of the shelves is extracted as one surface segment and 2 edge segments.
oin t Cloud Structure Graphs 141
Figure 5.44: Segmentation results for a LiDaR data set representing a portion of the Maltese Archipelago, consisting of both urban and agricultural terrain. Top left-hand side illustrates raw point cloud. The overlap (increased sample density) between successive aerial scans is clearly visible. Top right-hand side shows all segments produces by our segmentation process, whereas the bottom row left illustrates the edge segments and right illustrates the surface segments.
5.6
Discussion
The generation of 3D point clouds has become increasingly common in many areas of research. Given this huge amount of data, algorithms are required which are able to process, organise, cluster and extract important information about it in order to help in the post-processing effort. Our results have shown that the proposed segmentation method, PaRSe, is a feasible approach towards achieving this goal. For each of the point clouds used, the segmentation primitives proposed in this chapter have been able to represent some meaningful structure. In the case of the Mnajdra temple (§5.5.2), segmentation produces segments containing surface points from individual stones on the apse. An edge·complex segment is extracted which represents the contour of these stones. Given the complex nature of the point cloud, standard region-growing and shape fitting algorithms are not able to produce these segments. The Hal-Tarxien case study shows how small details in the site are identified in the segmentation process and represented using the segment primitives used (Figure 5.31). A graph query is used to identify segment patterns representing trees in the point cloud acquired at a University of Warwick green area. In this case, edge·complex segments are used to determine the location of trees. Segments resulting from the segmentation of a LiDaR point cloud include meaningful objects such as trees, houses, fields and streets. Similarly, on a smaller scale of two office environments, PaRSe computes a set partition whose elements represent a variety of objects which are typically found in an office including computer desktop, monitors, shelving, chairs, tables and other small objects. The automatic partitioning of the input point cloud into meaningful smaller segments helps in reducing the post-processing effort required to process the raw data.
With segmentation algorithms using RanSaC shape fitting, as the number of iterations/trials computed is limited, the solution (fitted planes in our case) obtained may not be optimal (Figure 5.28) and it may not even be one that fits the data in a good way. This is shown for a number of examples. In our case, since RanSaC fitting is done over segments which are previously established from a region-growing process and not the entire point cloud, this whole process is much more efficient (lines 13 and 14 of Algorithm 8) and the solution obtained for a complex site such as the Mnajdra temple is repeatable, i.e. given a number of runs of PaRSe and that the same input parameters are used, the same (or very similar) set partitions are produced fitting the raw data. Moreover, the
structure afforded by this repeatability, results in a higher level of abstraction over the data, which allows for the creation of query graphs which can be used to efficiently select different parts of a point cloud. For instance, all the trees in the point cloud of the Warwick university green area are segmented in a very similar way, which enables the query graph to identify all the trees.
The segmentation pipeline presented is efficient both in terms of memory and time complexities with data access patterns favouring parallelization. Region- growing may proceed concurrently (depending on the number of CPU cores) to produce a set partition of surface and edge segments. These segments are then analysed concurrently with the purpose of fitting plane primitives within them. An alternative task subdivision approach, since region-growing is generally less time consuming, is to adopt a consumer-producer scenario where region grown segments from a single thread are inserted in a pool for the rest of the threads to apply RanSaC plane fitting concurrently.