covariance matrix of local points. Detecting pylons play a key role to isolate individual spans. The previous works in the PL detection and modeling can be divided into two categories: 3D point-based approaches (Melzer and Briese 2004, Clode and Rottensteiner 2005, McLaughlin 2006, Vale and Gomes-Mota 2007, Jwa et al. 2009) and 2D image-based approaches (Yan et al. 2007, Li et al. 2010). Existing methods have demonstrated its successful performance, but in limited environments. For instance, McLaughlin (2006), Yan et al. (2007), and Jwa et al. (2009) reported approximately 72%, 90%, and 94% complete PL extraction rate in the point clouds and imagery data, respectively. The modeling errors mainly depend on scene complexity and data characteristics. In 2D images, the reported approaches might be sensitive to varying image gradients between PLs (voltage types) and their background (i.e., rural versus urban areas). In 3D point clouds, an irregularity in data distribution due to systematic/random errors and occlusions leads to failures in PL modeling completion.
3. METHODOLOGY Figure 1 schematically illustrates the overall workflow of the subsequent processes in which three major components, PL scene segmentation, pylon localization, and 3D object modeling are proposed in this study.
planarity amongst them are also considered as key properties to distinguish each others. We adopt Markov Random Field (MRF) to characterize unique (data attachment) and contextual properties of those features. In a graph model G = (V, E) comprised of vertices (V) and edges (E), the potential energy of a vertex is calculated by
(1)
where is the data energy term to describe the vertex and is the contextual energy term to describe the consistency with neighbouring vertex linked as a edge. and are the weighting parameters of and respectively (Baillard and Maître, 1999). When over all the vertices is minimized, the iterative optimization process is finished.
3.1.1 PL classification: A first step of the classification is to label point clouds into two groups: {linear, non-linear}. PL and pylon belong to linear features. LiDAR space is decomposed of a number of voxels (1.5m×1.5m×1.5m). Using all points within each voxel, linear candidate points are extracted, mainly using HT-like voting scheme which was proposed by Jwa et al. (2009) (Figure 2(b)). The linear candidate points are then converted into line segments by 3D line fitting based on RANdom SAmple Consensus (RANSAC) (Figure 2(c)). A graph model is generated on the basis of the line segments (i.e., vertices, V) and the connections between them (i.e., edges, E). Only locally connected edges are remained in the graph, i.e., if the distance between a pair of vertices comprising an edge is larger than a specified threshold (1.5m), the edge is discarded. To discriminate PL from others (mostly pylon), we investigate three data terms (height, line slope, and in-out cylinder) and one contextual term (line parallelism) in MRF. The weights of the terms are tuned by trial and error (0.25, 0.25, 0.2, and 0.3). Each of the energy terms has different potential functions depending on the states (i.e., PL or non-PL) of current vertex or neighbouring vertices. The potential energy functions are manually designed according to following bases.
height: PL is overhead, but part of pylon close to ground has low height.
line slope: PL is horizontal, while pylon is vertical except horizontal cross-arms.
in-out cylinder: is a ratio of point counts in two coaxial (inner of 0.15m radius and outer of 1m) cylinders created from a line segment. PL hardly has points in the space between inner and outer cylinder.
line parallelism: a single PL is generally parallel to others, but pylon shows random distribution.
Our classification problem is to find an optimal state of each line segment minimizing a global energy computed from the potential functions (Figure 2(d)). The state of the line segment having the highest energy is iteratively switched until the global energy converges. The member points of PL line segments are finally labelled into PL (Figure 2(h)).
3.1.2 Building classification: for each voxel, a plane segment is produced using the remaining points (non-linear) after PL classification. RANSAC is employed to find the best plane approximation (Figure 2(e)). We create locally connected graph consisting of vertices (plane segments) and edges (links of pairs of plane segments) as performed in Section 3.1.1. Two data terms (height and plane normal) and one contextual term (co-planarity) are examined for the separation of building and

vegetation which are dominant among the non-PL points.
Figure 1. Overall workflow for 3D reconstruction of objects including PL, pylon, and building in the PL corridor scene.
3.1 PL and building classification
Figure 2(a) presents a typical PL scene in urban cities which contains ground, PL, electric pylon, building, vegetation, vehicle, other low objects. The first classification step aims to label laser point clouds into linear and non-linear features, which are subsequently classified into PL and building objects. A key aspect in our classification is to consider PL as a set of linear features, while building as a set of planar features. Moreover, contextual properties such as co-linearity or co-
英语 ISPRS journal of photogrammetry and remote sensing 国际摄影测量和遥感学会... Mat 英语 Annals of geophysics 地球物理学纪事 英语 英语 Minerals & ...
英语 ISPRS journal of photogrammetry and remote 国际摄影测量和遥感学会志英语 ...(英文版) Technology Beijing, Mineral, 英语 Metallurgy, Mat Annals of ...
英语 ISPRS journal of photogrammetry and remote sensing 国际摄影测量和遥感学会... Mat 英语 Annals of geophysics 地球物理学纪事 英语 英语 Minerals & ...
ISPRS Journal of Photogrammetry and Remote Sensing 64 (2009) 472–481 ...Annals IX Simposio Brasileiro de Sensoriamento Remoto. INPE. Santos, Brazil....
ANNALS OF MATHEMATICS ANNALS OF THE NEW YORK ...ISPRS JOURNAL OF PHOTOGRAMMETRY AND REMOTE SENSING ...167 168 169 JOURNAL OF PROTEOME RESEARCH JOURNAL ...
Isprs journal of photogrammetry and remote sensing ...annals of geophysics 0.413 pure and applied ...文档贡献者 kj99102hsl 贡献于2012-02-28 ...
2012年SCI杂志中科院分区表(完整版)_其它_高等教育...ISPRS J PHOTOGRAMM IEEE POWER ENERGY M CEMENT ...ANNALS OF BIOMED APPLIED SOFT INFORMATION & ...
ISPRS J PHOTOGRAMM ISPRS JOURNAL OF PHOTOGRAMMETRY AND REMOTE SENSING 1.6740...ANNALS OF GLACIOLOGY 地学 0260-3055 4 0.8020 2012 B SOC GEOL FR ...
ISPRS JOURNAL OF ISPRS J PHOTOGRAMMETRY AND REMOTE...ANNALS OF GLACIOLOGY ANN GLACIOL AUST J EARTH ...2012 0.9837 1546 430 挪威 400 荷兰 347 法国 ...
ISPRS J PHOTOGRAMM ISPRS JOURNAL OF PHOTOGRAMMETRY AND REMOTE SENSING 0924-...ANNALS OF THE NEW YORK ACADEMY OF SCIENCES 0077-8923 PHILOS PHILOSOPHICAL T...
热门文档