APPLICABILITY ANALYSIS OF CLOTH SIMULATION FILTERING ALGORITHM FOR MOBILE LIDAR POINT CLOUD

Classifying the original point clouds into ground and non-ground points is a key step in LiDAR (light detection and ranging) data post-processing. Cloth simulation filtering (CSF) algorithm, which based on a physical process, has been validated to be an accurate, automatic and easy-to-use algorithm for airborne LiDAR point cloud. As a new technique of three-dimensional data collection, the mobile laser scanning (MLS) has been gradually applied in various fields, such as reconstruction of digital terrain models (DTM), 3D building modeling and forest inventory and management. Compared with airborne LiDAR point cloud, there are some different features (such as point density feature, distribution feature and complexity feature) for mobile LiDAR point cloud. Some filtering algorithms for airborne LiDAR data were directly used in mobile LiDAR point cloud, but it did not give satisfactory results. In this paper, we explore the ability of the CSF algorithm for mobile LiDAR point cloud. Three samples with different shape of the terrain are selected to test the performance of this algorithm, which respectively yields total errors of 0.44%, 0.77% and1.20%. Additionally, large area dataset is also tested to further validate the effectiveness of this algorithm, and results show that it can quickly and accurately separate point clouds into ground and non-ground points. In summary, this algorithm is efficient and reliable for mobile LiDAR point cloud.  Corresponding author. E-mail address: wumingz@bnu.edu.cn


INTRODUCTION
MLS technology can accurately and quickly acquire three-dimensional LiDAR (light detection and ranging) point cloud of earth surface.LiDAR point cloud filtering, which is to separate point clouds into ground and non-ground points, is an essential step in post-processing.
Many ground filtering algorithms have been proposed during previous decades, and these algorithms can be mainly divided into three categories (Zhang et al., 2016): slope-based methods (Sithole, 2001;Susaki, 2012;Vosselman, 2000), mathematical morphology-based methods (Chen et al., 2007;Li, 2013;Li et al., 2014;Zhang et al., 2003) and surface-based methods (Nie et al., 2017;Pfeifer et al., 1999;Zhao et al., 2016).The In recent years, cloth simulation filtering (CSF) algorithm is proved to be an accurate, automatic and easy-to-use algorithm for airborne LiDAR point cloud.Specifically, the accuracy of this algorithm is comparable with most of the state-of-the-art ground filtering algorithms, and its parameters are few and are easily set by the users without much experience.In addition, this algorithm has been developed a CloudCompare plugin.
Due to the different speed of travel, trajectory and scan distance of the laser scanning system, there are some differences between airborne and mobile LiDAR point clouds (Table 1) The remainder of the paper is organized as follows.The principle of the CSF algorithm is described in Section 2. Next, the experiments are performed and the results are analyzed in Section 3. Finally, the conclusion is given in Section 4.

METHOD
The method is based on the simulation of a simple physical process.Imagine a piece of soft enough cloth placed above the terrain, and then the cloth falls under the action of gravity.The final shape of the cloth is the DSM (digital surface model).In contrast, if the surface is turned upside down, and then a cloth with rigidness falls under the action of gravity, the final shape of the cloth is the DTM.
Figure 1.Overview of the CSF algorithm.
To simulate this physical process, CSF algorithm utilizes a cloth simulation technique to separate point clouds into ground and non-ground points (Zhang et al., 2016).Figure 1 illustrates the overview of this algorithm.The procedure of the algorithm is shown as follows (Figure 2):

EXPERIMENTS AND RESULTS
Three case studies are illustrated to assess the performance of The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLII-3, 2018 ISPRS TC III Mid-term Symposium "Developments, Technologies and Applications in Remote Sensing", 7-10 May, Beijing, China the CSF algorithm.As shown in Figure 3a, various non-ground objects (such as buildings, trees and poles) exist in the study area.In particular, the features of the topography for three cases are significantly different, including flat terrain, gentle slope terrain and high slope terrain.The reference datasets were obtained by manually classification, and each point was labeled as ground or non-ground points.
To verify the effectiveness of the CSF algorithm in large area, we also tested a large area dataset, which mainly include 1000m road and area on both sides of road (the size is  In Figure 3b, it can be seen the CSF algorithm can filter out most of non-ground points from the original point cloud with the different features of terrain, and the terrain characteristics can be effectively remained as well.
For the large area dataset, the algorithm costed about 1.2 min to finish the experiments.By the visual inspection, we can see that this algorithm can successfully filter out non-ground points, as shown in Figure 4b.
Moreover, qualitative evaluation of the CSF algorithm was implemented by using reference samples.In this research, type I, type II and total errors were utilized to quantitatively assess the performance of this algorithm.Specifically, type I error is the rate of ground points misclassified as non-ground points, type II error is the rate of non-ground points misclassified as ground points, and the total error is the rate of misclassified points, which equation is presented as follows: Table 3 shows the accuracy assessment of CSF algorithm, and Figure 5 shows the corresponding spatial distributions of the type I and type II errors, which indicates that this algorithm have high precision for all reference samples, and the total error is less than 1.20% for all the results.The main reason of high precision is that the simulated cloth can be directly treated as the final generated DTM for some circumstances, which avoids the interpolation of ground points, and can also recover areas of missing data.In addition, the distribution complexity of non-ground objects seldom influences the terrain approximation process (Zhang et al., 2016).

Samples
aforementioned ground filtering algorithms has proven to be successful for airborne LiDAR point cloud.However, these algorithms commonly have the following problems: (1) parameters setting are complicated; (2) filtering results are usually unreliable in complex areas; (3) most of them are not open source.The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLII-3, 2018 ISPRS TC III Mid-term Symposium "Developments, Technologies and Applications in Remote Sensing", 7-10 May, Beijing, China , in which distribution feature plays an important role in the filtering result.There is much missing in mobile LiDAR point cloud due to occlusion of objects, which lead that the distribution feature is very unevenly.As a result, aforementioned filtering algorithm did not give satisfactory results for mobile LiDAR point cloud.Differences between airborne and mobile LiDAR point cloud In this paper, we explored the performance of the CSF algorithm for mobile LiDAR point cloud.Three reference samples with the different characteristics of the terrain were tested.

Figure 2 .
Figure 2.The flowchart of the CSF algorithm

Figure 3 .
Figure 3.Results of each sample: (a) original datasets colored by height; (b) the ground points extracted from the CSF algorithm.

Figure 4 .
Figure 4.Result of large area dataset: (a) original dataset colored by height; (b) the ground points extracted from the CSF algorithm.
a represents the number of ground points misclassified as non-ground points, b represents the number of non-ground points misclassified as ground points, and c and d represent the total number of ground and non-ground points, respectively.

Figure 5 .
Figure 5.The spatial distributions of the type I and type II errors

Table 3 .
Three types of error statistics