STREET-SCENE TREE SEGMENTATION FROM MOBILE LASER SCANNING DATA

Our work addresses the problem of extracting trees from mobile laser scanning data. The work is a two step-wise strategy, including terrain point removal and tree segmentation. First, a voxel-based upward growing filtering is proposed to remove terrain points from the mobile laser scanning data. Then, a tree segmentation is presented to extract individual trees via a Euclidean distance clustering approach and Voxel-based Normalized Cut (VNCut) segmentation approach. A road section data acquired by a RIEGL VMX-450 system are selected for evaluating the proposed tree segmentation method. Qualitative analysis shows that our algorithm achieves a good performance. * Corresponding author


INTRODUCTION
Mobile laser scanning (MLS), a widely used technology since year 2003 when the first MLS system was developed, has attracted much attention for mainly transportation-related surveys (Jacobs, 2005;Toth, 2009).It is a data revolution.With an MLS system, mobile mapping engineers can drive on a highway, rural road, and railroad, or along the shoreline of a river or lake.Along the way, the system captures trees, bridges, streetlights, buildings, power lines, other street-scene small objects (e.g.cracks, road markings).The collected data are a totally immersive three-dimensional (3D) view of the objects and surroundings (Rybka, 2011).
An MLS system, integrated with a navigation solution, laser scanner(s), high-resolution camera(s), and even more powerful computer systems, has been commercially available for several years (Stauth and Olsen, 2013).MLS systems are capable of working on all-day and all-weather conditions without the consideration of environmental illuminations.In addition, the acquired 3D point cloud data are with real-world coordinates, thereby providing accurate 3D geospatial information of roadways.Thus, studies on MLS as a reliable and cost effective alternative, is worthwhile for carrying out road-scene object inspections along route corridors.
In the case of urban areas, the detection and quantification of vegetation is important for road-scene feature inventory (for example, in safety studies and noise modelling) and environmental and ecological analysis.As an important component of urban vegetation, street trees play an important role in maintenance of environmental quality, aesthetic beauty of urban landscape, and social service for inhabitants.Most developed tree detection methods were applied to terrestrial laser scanning (TLS) point clouds.These methods range from 3D Hough transform, mathematical morphology to skeleton approaches by fitting free-forms.
Based on research achievement, Wu et al., (2013) presented a new step-wise Voxel-based Marked Neighbourhood Searching (VMNS) method for efficiently identifying street trees and deriving their morphological parameters from MLS data.The presented tree detection method included voxelization, calculating voxel values, searching and marking neighbourhoods, extracting potential trees, deriving morphological parameters, and eliminating pole-like objects.Shen et al. (2008) adopted the minimum spanning tree method for extracting objects (including trees) from scattered MLS points.Rutzinger et al., (2010) segmented initially point clouds into planar regions using a 3D Hough transform and surface growing algorithm.A connected component analysis was further used to merge the remaining small segments.Some object statistics were calculated to obtain tree clusters by removing nonvegetation objects and temporary objects, such as cars and pedestrians.Finally, the classified tree points were thinned by deriving the vertices forming the 3D alpha shape of point clouds.A sphere with a certain radius (alpha value) was passed iteratively from the outside towards the points.Finally, the values of model parameters (describing only the outer shape of the tree.)were derived from the extracted single tree point clouds.Saarinen et al., (2013) used multi-temporal MLS data sets to detect the change of riverine vegetation.The proposed method first placed a 2 m × 2 m grid over the study area, and derived metrics describing vegetation density and height to predict the variables in the nearest-neighbour (NN) estimations.According to the training data obtained from aerial images, the vegetation cover type was classified into the following four classes: bare terrain, field layer, shrub layer, and canopy layer.The workflow combining detection and modelling was applicable on single trees scanned by TLS and also on larger areas acquired by MLS.The overall shape of the tree originates from the length of the primary branches and their position on the trunk.This paper propose a two step-wise tree segmentation method, which includes (1) road point removal based on voxel-based upward growing filtering, and (2) tree segmentation based on a voxel-based normalized cut (VNCut) segmentation method.The remainder of this paper is organized as follows.Section 2 describes the used MLS data.Section 3 introduces the proposed tree segmentation method.Section 4 demonstrates the experiment results and discussion.Finally, concluding remarks are given.

MOBILE LASER SCANNING TEST DATA
The data were acquired on 23 April 2012 by a RIEGL VMX-450 system.The surveyed area is in a tropical urban environment, located in the City of Xiamen, southeast China.A complete survey was carried out, once in a forward direction and once in a reverse direction, on Huandao Road from Xiamen University to International Conference and Exhibition Center (ICEC), at a driving speed of approximately 50 km/h.This twodirectional-four-lane Huandao road is separated by a median.This is a typical subtropical urban road area with high buildings, dense vegetation, and traffic signposts on both sides of the road.
The surveyed road was quite busy; the average driving speed ranged from 30 ~50 km/hour.The buildings (e.g., high-rise residential apartments and commercial buildings) are located along this typical urban road.This survey kept 30 m and 30 km/hour for target distance and ground speed, respectively.The point density for MLS data strongly relies on the nominal distance to the target where the point spacing is measured as well as the incidence angle.Accordingly, at the vehicle speed of 30 km/hour, the values of point density are estimated as 286.44 points/ m 2 .Figure 1 shows a road section of the collected Huandao-Road points.

METHODOLOGY
The proposed method is two step-wise strategy: (1) the separation of terrain points and non-terrain points via a voxelbased upward growing filtering, and (2) individual tree segmentation based on a Euclidean distance clustering approach and a voxel-based Normalized Cut appraoch.

Terrain Point Removal
MLS systems collect very dense data close to the scanner path and less dense data farther away from the scanner path.Points belonging to road surface account for a great portion of the collected MLS data.Thus, the established filtering algorithms are unsuitable for filtering non-terrain points from MLS data.To improve computational efficiency, we develop a rapid and effective method, namely voxel-based upward growing filtering method, for removing terrain points.This method is implemented as follows:

Tree Segmentation
To rapidly group non-terrain points into separated clusters representing individual objects, we use a Euclidean distance clustering approach, which clusters discrete points based on their Euclidean distances.Theoretically, an un-clustered point is grouped into a specific cluster if and only if the shortest Euclidean distance to the points in this cluster lies below a predefined threshold, dt.Otherwise, a new cluster is created to contain this point.Such a Euclidean distance clustering approach starts at an un-clustered non-terrain point and then iteratively processes each un-clustered point until all non-terrain points are grouped into specific clusters.Through Euclidean distance clustering, most separated objects are successfully segmented.However, by using the Euclidean distance clustering approach, some adjacent or overlapped objects cannot be well segmented.
Before tree classification, individual trees must be separated from the clusters that contain multiple objects.In our previous study (Yu et al., 2015), we propose a voxel-based normalized cut (VNCut) segmentation method, which effectively segments point cloud clusters, containing multiple adjacent or overlapped objects, into separated objects.Thus, in this letter, we use the VNCut segmentation method to segment individual trees.
First, a cluster containing multiple overlapped objects is voxelized using the octree partition strategy with a voxel resolution ws.Then, the generated voxels are organized into a complete weighted graph G = {V, E}, where the vertices V are formed by the voxels, and the edges E are connected between each pair of voxels.The weights on the edges function to measure the similarities between the two connected voxels.Such a weight is computed based on the geometric features of the associated voxels as: where wij = the weight on the edge connecting voxels i and j pi = (xi, yi, zi) = the centroid of voxels i pj = (xj, yj, zj) = the centroid of voxels j The centroid of i is computed as: where mi = the number of points in voxel i; i k p = a point in voxel i.
Thus, if the horizontal distance between two voxels exceeds dH, the weight on the edge connecting these two voxels is set to zero.In the standard normalized cut segmentation method (Shi and Malik, 2000), the cost function that partitions graph G into two disjoint voxel groups A and B by maximizing the similarity within each voxel group and minimizing the similarity between two voxel groups is defined as: By using the VNCut segmentation method, the clusters that contain more than one object are effectively segmented into individual objects.After segmentation, prior knowledge (e.g.crown size) is used to filter out those non-tree objects such as light poles.Figure 3 shows a visual example of the detected individual trees.

RESULTS AND DISCUSSION
Figure 4 (a) shows a road section selected from the data (shown in Figure 1) for evaluating the proposed tree segmentation performance.This road section is about 60 m long.Some short trees (Sago cycas) are regularly planted in the middle of the road.Along two sides of the road, some trees are very dense and even overlapped with each other.

CONCLUSION
This paper has presented a two step-wise tree segmentation method from MLS data.To efficiently identify trees, the voxelbased upward growing filtering method was first presented to classify MLS data into terrain points and non-terrain points.
Next, based on the classified non-terrain points, the Euclidean distance clustering was performed to obtain a set of clusters and then the voxel-based normalized cut method was used to obtain individual trees from these clusters.To evaluate the performance of the proposed tree segmentation, a road section was selected from the MLS data acquired by a RIEGL VMX-450 system.The visual inspection of the experimental results demonstrated the feasibility of the proposed method in segmenting individual trees from MLS data.

Figure 1 .
Figure 1.A road section of MLS data collected along Huandao Road in Xiamen, China.

1)
Figure 2. (a) Gridding raw points of a road section into data blocks in the XY plane, (b) octree partition structure for block Dj, and (c) nine upper neighbors (N1 to N9) for each voxel vi.
z coordinate of the centroids of voxels j 2 H  = the variances of the horizontal distributions 2 V  = the variances of the vertical distributions dH = a distance threshold restraining the maximum valid horizontal distance between two voxels Figure 3. (a) Non-terrain points after terrain removal, (b) detected individual trees after VNCut.

First
Figure 4.A test data selected from Figure 1.(a) The raw point cloud, (b) the classified terrain points and non-terrain point, (c) the classified non-terrain points, and (d) the individual tree segmentation results.