grouped as one segment. A well-known algorithm based on the model fitting approach is the RANdom SAmple Consensus (RANSAC) proposed by Fischer and Bolles, (1981). This algorithm was applied for automatic processing ofpointcloud with the aim of3D building modeling (Tarsh-Kurdi et al., 2007). The main shortcoming of model fitting approaches is their inefficiency and spurious segmentation results when dealing with different pointcloud sources (Filin, 2002). The segmentation method based on clustering of attributes is a robust approach for the identification of homogenous patterns in the data. This method mainly comprises two processes: attribute computation, and clustering the data based on the computed attributes. Since this method is highly dependent on the quality of derived attributes, they should be computed precisely to produce the best separation among different classes. These techniques generate a voting scheme in the attribute space which is constructed using an accumulator array. The dimension of this accumulator array is dependent on the number of the utilized attributes for clustering. Vosselman and Dijkman (2001) used the principal of the Hough transform (Hough, 1962) for segmentationof planar surfaces in a 3Dlaserpointcloud. In this method, each laserpoint defines a plane in the 3D attribute space. So, the laser points on the same planar surface will intersect at the position in the attribute space that corresponds to the slopes and distance of the planar surface. Filin and Pfeifer (2006) introduced a segmentation method based on the normal vectors derived using a slope adaptive neighbourhood. They used the slopes of the normal vector in the X and Y directions and height difference between the point and its neighbourhood as the clustering attributes. This height difference attribute was also used to guarantee the distinction between parallel planes, which share the same normal vector slopes. Biosca and Lerma (2008) suggested a fuzzy clustering approach in combination with a similarity-based cluster merging for segmentationof a terrestrial laserpointcloud. Kim et al. (2007) proposed a method for segmentationof planar International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXVIII-5/W12, 2011
A further option to incorporate more regional context into the classification process are CRF with higher order potentials. Na- jafi et al. (2014) set up a non-associative variant of this approach for point clouds. They first performed a segmentation and then applied the CRF to the segments. Overlapping segments in 2D were considered by a higher order potential. The authors addi- tionally modelled the object class relation in the vertical direc- tion with a pattern based potential. This is useful for terrestrial scans, but in the airborne case the derived features for the vertical are not very expressive due to missing point data for example on fac¸ades. Although higher order potentials are becoming more and more important, it is still difficult to apply them to point clouds (in particular for point-based classification) due to the extensive computational costs. Inference for such models is a challeng- ing task and up to now only very few nodes can be combined to form a higher order clique for non-associative interaction models which currently restricts the expressive power of this framework. In the case of Najafi et al. (2014) only up to six segments were combined to one higher order clique to deal with this problem. Xiong et al. (2011) showed how point-based and region-based classification of LiDAR data can interact in a pairwise CRF. They proposed a hierarchical sequence of relatively simple classifiers applied to segments and points. Starting either with an indepen- dent classification of points or segments, in subsequent steps the output of the previous step is used to define context features that help to improve the classification results. In each classification stage, the results of the previous stage are taken as input, and, unlike with a single CRF, it is not guaranteed that a global op- timum is reached (Boykov and Jolly, 2001; Kumar and Hebert, 2006). Luo and Sohn (2014) applied two asymmetric (pairwise) CRF for short range and long range interactions separately on terrestrial scan line profiles and evaluated the experiment on a terrestrial laserpointcloud. The final label for each laserpoint is determined by finding the maximum product of both CRF poste- riors. While the short range CRF has a smoothing effect on the point labels, a CRF for long range interaction models the struc- ture of the scene. It was found that the introduction of the long ranges context was able to eliminate some misclassification er-
The idea of optimization (Eq. 2) is to determine the filtering degree such that a quasi-optimal relationship is formed between the considered forest attribute and an attribute characterizing the filtrations. Vauhkonen et al. (2014) used the modeled canopy- and above-ground biomasses for this purpose, but found the resulting canopy volumes to also have a strong allometric relationship with all other forest attributes considered. As a variate of the tree diameter, G, on the other hand, is directly measured from the trees and also the most typical forest attribute obtained in aggregate-level forest inventories. The use of G was thus justified by the practical applicability of the proposed approach, but also by the unknown form of the “real” allometric relationship between the canopy volume and forest attributes of interest (see Discussion by Vauhkonen et al. 2014). Whether the relationship formed by other variables was better justified, both G and the tetrahedral volume currently used for characterizing the simplicial complexes could be substituted in Eq. 2 by these variables.
Laser instrument is connected to a portable PC where a software able to set the acquisition, to archive and pre-process the collected data is installed. Based on the setting parameters,, mainly related to the area of scanning and to the requested spatial resolution, the instrument acquires the 3Dpointcloud in a time span and at a density according to the technical characteristics of the adopted instrument. The acquisition software creates a project for each dataset, which contains the raw data for each scans, the digital images, ancillary information, etc. It also allows a pre-processing step that firstly perform the removal of noise and blunder data filtering. Successively co-registration of different scans, creation of a solid models and integration with other information (mapping of digital images) can be performed. More specifically, the pre- processing step includes the filtering, alignment, georeferencing, segmentation, modeling and texture, that are shown in Figure 3.
traffic signs (Wen et al., 2016; Yu et al., 2016b), street light poles (Yu et al., 2015b; Wu et al., 2016), and parking cars (Yu et al., 2016c). Traditionally, digital images were the preferred source of road crack surveying which can derive road cracks in millimetre width. (e.g., Xu and Huang, 2003). However, image qualities often depend on weather, traffic, and image photogrammetry techniques (Nguyen et al., 2011). Pavement crack detection using MLS point clouds becomes a new research topic recently. The idea was basically to make the 3D information of MLS data aid in pavement distress analysis. However, the large volume, mixed-density and irregular distribution of3D MLS points make road surface’s object extraction very challenging and difficult. Our efforts have been placed on development of novel algorithms and software tools for accurate and robust object extraction. To achieve this goal, current MLS road crack extraction methods rely on converting 3D pavement information into 2D image-based data and crack delineation is performed using image processing approaches. Previous studies on pavement cracks detection can achieve precision in centimetre using direct detection from 3Dpoint clouds. Guan et al. (2014) proposed an iterative tensor voting algorithm to detect road cracks from geo-referenced images. The iterative tensor voting based crack detection achieved completeness as 96% and correctness as 85%. It could well detect road cracks with width larger than 2 cm. Tsai and Li (2012) implemented a dynamic-optimization-based crack segmentation method on high-resolution 3D continuous transverse pavement profiles. The continuous transverse pavement profiles acquired from a MLS system could aid in detecting road cracks with widths greater than 2 mm. It enabled crack detection under low intensity contrast and lighting conditions with above 95% score of precision. The installation of both downward camera and the 360 degrees laser scanning could be available to traffic safety maintenance.
vertically configured laser sensors covering a wide enough verti- cal FOV (depending on the number of sensors). It rotates around the vertical axis such that it generates a panoramic view of the surroundings. The rotation frequency ranges from 6 to 15 Hz. Same as other types oflaser scanners, the Velodyne range data are directly recorded in 3D in the form of3Dpoint clouds. The measurement distance ranges from 2 m to 100 m, and the range accuracy is about 2 cm. The Velodyne scanner constantly scans the full surroundings hence it is an ideal technique for MODAT, especially when the area of interest is located around the sensor (Moosmann and Stiller, 2013). It can be mounted on a mobile mapping system (MMS), together with optical cameras, for the purposes of environment perception, and simultaneous localiza- tion and mapping (SLAM) (Moosmann and Stiller, 2011). A popular MODAT method is tracking-by-detection, where the moving objects are detected first in each frame, then the trajec- tories are reconstructed by associating plausible candidates (An- driluka et al., 2008, Wu and Nevatia, 2007). Objects are typically detected by extracting discriminative features from pixels or seg- ments generated from segmentation methods, then they are classi- fied into objects of interest. However, the detection results can be affected by many factors, such as occlusion, miss-classification. To avoid miss detections using classification, some try to track generic objects without knowing the specific classes, which, how- ever, can have limited applications (Kaestner et al., 2012). In both cases, the detection accuracy will limit the overall tracking per- formance.
Computing the visible part of a 3D object is a vital problem in computer graphics, computer vision, robotics, GIS and photogrammetry. Usually the visibility should be accomplished in an automated way from a certain viewpoint or camera. Currently, the point clouds can be produced either by using laser scanning or dense image matching which is widely used for 3D acquisition, representation and reconstruction. These point clouds are either sparse or dense of millions points. However, a problem arises when viewing a pointcloud as shown in Figure 1 where the objects looking direction cannot be identified (Katz et al., 2007). This necessitate to use the visibility testing and to discard the occluded points to properly view the object points.
Over the past decade, large-scale photogrammetric products have been extensively used for the geometric documentation of cultural heritage monuments, as they combine metric information with the qualities ofan image document. Additionally, the rising technology of terrestrial laser scanning has enabled the easier and faster production of accurate digital surface models (DSM), which have in turn contributed to the documentation of heavily textured monuments. However, due to the required accuracy of control points, the photogrammetric methods are always applied in combination with surveying measurements and hence are dependent on them. Along this line of thought, this paper explores the possibility of limiting the surveying measurements and the field work necessary for the production of large-scale photogrammetric products and proposes an alternative method on the basis of which the necessary control points instead of being measured with surveying procedures are chosen from a dense and accurate pointcloud. Using this pointcloud also as a surface model, the only field work necessary is the scanning of the object and image acquisition, which need not be subject to strict planning. To evaluate the proposed method an algorithm and the complementary interface were produced that allow the parallel manipulation of3Dpoint clouds and images and through which single image procedures take place. The paper concludes by presenting the results of a case study in the ancient temple of Hephaestus in Athens and by providing a set of guidelines for implementing effectively the method.
Segmentation in 3Dpoint clouds obtained from laser scanners is not trivial, because the three dimensional point data are usually incomplete, sparsely distributed, and unorganized, also there is no prior knowledge about the statistical distribution of the points, and the densities of points vary with the point distribution. Many methods have been developed to improve the quality of segmen- tation in 3Dpoint clouds that can be classified into three main categories: edge/border based, region growing based and hybrid. The edge/border based methods attempt to detect discontinuities in the surfaces that form the closed boundaries, and then points are grouped within the identified boundaries and connected edges. These methods usually apply on the depth map where the edges are defined as the points where the changes of the local surface properties exceed a given threshold. The local surface properties mostly used are surface normals, gradients, principal curvatures, or higher order derivatives (Sappa and Devy, 2001, Wani and Arabnia, 2003). However, due to noise caused by laser scanner- s themselves or spatially uneven point distributions in 3D space,
This paper compares two generic approaches for the reconstruction of buildings. Synthesized and real oblique and vertical aerial imagery is transformed on the one hand into a dense photogrammetric 3Dpointcloud and on the other hand into photogrammetric 2.5D surface models depicting a scene from different cardinal directions. One approach evaluates the 3Dpointcloud statistically in order to extract the hull of structures, while the other approach makes use of salient line segments in 2.5D surface models, so that the hull of3D structures can be recovered. With orders of magnitudes more analyzed 3D points, the pointcloud based approach is an order of magnitude more accurate for the synthetic dataset compared to the lower dimensioned, but therefor orders of magnitude faster, image processing based approach. For real world data the difference in accuracy between both approaches is not significant anymore. In both cases the reconstructed polyhedra supply information about their inherent semantic and can be used for subsequent and more differentiated semantic annotations through exploitation of texture information.
In this paper we propose a new method for automatically approx- imating the size parameters and structures of trees from point clouds. The basic assumptions of the method are that the pointcloud is a sample of a surface in the 3D-space and the surface, i.e. the tree, can be locally approximated with cylinders. Other a pri- ori assumptions about the data and structure of trees are used as well. The basis of the method is a local approach where the pointcloud is covered with small neighborhoods which conform to the surface. Then these neighborhoods are geometrically character- ized and, based on these characterizations, the neighborhoods are classified into trunk, branch, and other points. Finally, cylinders are fitted to proper subsets to approximate the size. Notice that voxel spaces are not used, although partitions of the pointcloud into cubical cells are used to produce the coverings quickly.
This paper presents a system that uses volumetric sensory data to detect small obstacles in level crossings. For this purpose it relies on the use of dense 3Dpoint clouds produced by a tilting 2D laser scanner. This way the system is not as dependent on lighting conditions and object’s appearance as it would if relying on binocular vision, and it is also not as dependent on ground planarity as are solutions based on static 2D laser scanners. The system is composed of a software layer that processes the 3Dpointcloud preceded by an acquisition hardware layer. Obstacle detection is carried out exploiting the assumption that a quasiplanar dominant ground plane exists. A background model of the level crossing is learned from a set of training point clouds. Obstacles are the objects that show up in the point clouds captured online that are inexistent in the training dataset. To avoid cumber- some and time-consuming manual on-site calibration, the proposed system includes an automatic calibration process that estimates the pose of both level crossing and railway with respect to the laser scanner.
The currently existing mobile mapping systems equipped with active 3D sensors allow to acquire the environment with high sampling rates at high vehicle velocities. While providing an effective solution for environment sensing over large scale distances, such acquisi- tion provides only a discrete representation of the geometry. Thus, a continuous map of the underlying surface must be built. Mobile acquisition introduces several constraints for the state-of-the-art surface reconstruction algorithms. Smoothing becomes a difficult task for recovering sharp depth features while avoiding mesh shrinkage. In addition, interpolation-based techniques are not suitable for noisy datasets acquired by Mobile Laser Scanning (MLS) systems. Furthermore, scalability is a major concern for enabling real-time rendering over large scale distances while preserving geometric details. This paper presents a fully automatic ground surface recon- struction framework capable to deal with the aforementioned constraints. The proposed method exploits the quasi-flat geometry of the ground throughout a morphological segmentation algorithm. Then, a planar Delaunay triangulation is applied in order to reconstruct the ground surface. A smoothing procedure eliminates high frequency peaks, while preserving geometric details in order to provide a regular ground surface. Finally, a decimation step is applied in order to cope with scalability constraints over large scale distances. Experimental results on real data acquired in large urban environments are presented and a performance evaluation with respect to ground truth measurements demonstrate the effectiveness of our method.
Automatic detection of individual building and recognition of its distinct sub-elements from remote sensing data are crucial for many applications including 3D building modelling, building level damage assessment and other urban related studies (Dong and Shan, 2013; Sun and Salvaggio, 2013). Generally, the buildings and its elements possess unique geometric characteristics. Hence, the 3D geometric features are being used as the fundamental information in building detection and categorisation of its sub-elements (Rottensteiner et al., 2014; Xiong et al., 2013). 3Dpoint clouds are well suited to infer the geometric characteristics of the objects. Particularly, the multi-view airborne oblique images are a suitable source to generate 3D points cloud for building analysis as they can provide information of both the roofs and facades of the building (Liu and Guo, 2014). Unmanned Aerial Vehicles (UAVs) are attractive platforms which can capture the images with suitable characteristics such as multi-view, high overlap and very high resolution to generate very dense 3Dpointcloud in minimal time and cost (Colomina and Molina, 2014). Generally, the building detection process from 3Dpoint clouds has been carried out through identifying planar segments as most elements of general buildings are planar surfaces (Dorninger and Pfeifer, 2008). Planar segments with its geometric features could help to detect and delineate buildings in the scene. However, an accurate segmentationof individual elements of the building is not always feasible, especially with the geometric features from image-based 3Dpointcloud. This is
Automated calibration of LIDAR systems has been an active field of research and development over the last years. Traditional calibration approaches rely on manual extraction of geometric features in the laser data and require time-intensive input of a trained operator. Recently, new methodologies evolved using automatic extraction of linear features and planar information to minimize systematic errors in LIDAR strips. This paper presents a new methodology of LIDAR calibration using automatically reconstructed planar features. The calibration approach presented herein integrates the physical sensor model and raw laser measurements and allows for refined calibration of internal system parameters. The new methodology is tested and compared with a traditional approach based on manual boresighting using a typical survey mission. Optech’s software suite LMS, which is the first commercial implementation of this functionality, was used to process the data and to derive means of quality assessment. Different methods of reconstructing automatically extracted geometric features are presented and discussed in the context of their contribution to the calibration process. The final results are compared numerically and through graphic quality check.
(3) where, M = N 2 + 1 . This expression clearly indicates that the beamwidth ofan array was inversely proportional to frequency. It implies that an increase in either the number of elements or interelement spacing results in a decrease in the beamwidth as well.
The experiments are performed on a 6DOF Zebra Zero ma- nipulator (IMI Inc.). A KPD-50 CCD camera (Hitachi Ltd.), with a lens of focal length f = 6.0 [mm], is mounted in front of the robot (see Figure 9 for a camera pointof view). The extracted visual features are the image coordinates of a white sphere centroid located at the robot wrist and its im- age projected surface. The images of 640 × 480 [pixel] are acquired using a Meteor frame-grabber (Matrox Ltd.) at 30 frames per second (FPS) with 256 grey levels. The image processing is performed on a 50 × 50 [pixel] sub-window, in order to guarantee that the sphere remains within the sub- window. The first estimations of the white sphere coordinates and area are performed off-line using a Graphical User In- terface (Figure 9), named VServo, developed in Tcl/Tk lan- guage. During task execution, features (centroid and area) are computed using the image moments algorithm (Haralick and Shapiro, 1993).
Future research in MRI segmentation should strive toward improving the accuracy, precision, and computation speed of the segmentation algorithms, while reducing the amount of manual interactions needed. This is particularly important as MR imaging is becoming a routine diagnostic procedure in clinical practice. It is also important that any practical segmentation algorithm should deal with 3D volume segmentation instead of 2D slice by slice segmentation, since MRI data is 3D in nature. Volume segmentation ensures continuity of the 3D boundaries of the segmented images whereas slice by slice segmentation does not guarantee continuation of the boundaries of the tissue regions between slices.
We can generate a two-dimensional map over the θh plane with a projection of the colors as seen in every direction captured by the cameras (Figure 5.5). The next step is to apply a 2D transform to this map, but we need to cope with its sparsity. For so, one may divide the cam- era directions (θh) plane into sub-regions of the same area, as depicted in Figure 5.7. Each area is a quantized description of the camera direction. We may further divide each sub-region sev- eral times until attaining the desired precision. The smaller the sub-region, the more precise the camera position is represented. The number of divisions should be chosen to make the position information for two or more cameras not to fall within the same sub-region. In Figure 5.7, after dividing the plane, several sub-regions remain unoccupied. This representation is similar to vox- elized point clouds in the 3D space. Therefore, we apply RAHT  to the colors associated with each camera (sub-region), through a 2D quad-tree decomposition rather than the 3D octree.
Rauch and Hendrickson (2004) classify securitized loans as ratio (statement) loans. Since credit scoring may be inefficient (BAAS; SCHROOTEN, 2006), securitized lending covered by liquid collateral (BONFIM, 2005; KIMBER, 2004) is more effective in terms of risk reduction and therefore imply low-risk credit contracts (MESTER, 1997), decreasing thus adverse selection. In fact, securitized loans may represent a strategy for lenders to acquire collateral which is valuable for them, as Ray (1998) suggests that lenders may grant credit with that objective. The liquidity of receivables is usually not very risky because suppliers have an advantage in collecting information about their non-financial customer’s, in accessing their credit worthiness and in controlling their actions; this informative advantage allows suppliers do discriminate between their good and bad customers better than banks (WILNER, 2000).