ISPRS Commission III, Vol.34, Part 3A „Photogrammetric Computer Vision‘, Graz, 2002 
  
FILTERING STRATEGY: WORKING TOWARDS RELIABILITY 
George Sithole 
Department of Geodesy, Faculty of Civil Engineering and Geosciences 
Delft University of Technology 
The Netherlands 
g.sithole@citg.tudelft.nl 
Commission III, Working Group 3 
KEY WORDS: laser scanning, LIDAR, DEM/DTM, classification, filtering 
ABSTRACT 
The filtering of a laser scanner point-cloud to abstract the bald earth has been an ongoing research topic in laser altimetry. To date a 
number of filters have been devised for extracting DEMs from laser point-clouds. The measure of the performance of these filters is 
often based on tests against some reference data (rms, ratio of misclassifications vs. correct classifications, etc.) obtained by 
photogrammetric measurement or other means. However, measures based on such tests are only global indicators of how the filters 
may perform. Therefore, when applied to real life applications, based on such measures it is not possible to say with certainty how 
well a filter has performed. This uncertainty suggests that a method be devised to identify in a point-cloud those regions where a 
filter may have difficulty classifying points. This done other sources of information can be gathered to clarify the status of points in 
(difficult) regions. This fits in with the thinking that external sources of data, such as imagery, maps have be used in the filtering of 
laser scanner point-clouds. However, devising a method as suggested above requires that the reasons for the misclassification of 
points be first identified. When filtering a point-cloud based on spatial information alone, misclassification arises from three sources, 
(1) the nature and arrangement of objects and the terrain in a landscape (e.g. terrain, buildings, vegetation, etc.,) (2) the 
characteristics of the data (resolution, outliers, data gaps, etc.,) and (3) the implementation of filters. In this paper, the first two 
reasons for misclassification are outlined because they are common to all filtering problems, and an initial attempt at developing a 
  
method for identifying regions in a point-cloud where a filter may have trouble in classifying points is described. 
1 INTRODUCTION 
The filtering of laser scanner point-clouds to abstract the bald 
earth has been an ongoing research topic in laser altimetry. 
Filtering in the context of this paper is understood to mean the 
removal from a laser scanner point-cloud of those points that do 
not belong to the terrain. To date various filters have been 
developed, e.g., (Kraus & Pfeipfer 1998), (Axelsson 1999), 
(Petzold et. al. 1999), (Elmqvist 2001), (Vosselman & Maas 
2001), (Haugerud et. al. 2001). Additional to these are 
proprietary filters (being used in practice), whose algorithms are 
not known. The performance of filters is currently measured by 
comparison of filter results with some reference data (rms, ratio 
of misclassifications vs. correct classifications, etc.,). However, 
there are three problems with performance measures based on 
reference data: 
* They are global indicators. The performance of filters is 
unified into one single measure. Unfortunately, experience 
has shown that this masks localized filtering errors (which 
can sometimes be large, although few in number). 
e The performance measures are indicative of filter 
performance only in areas that have characteristics similar 
to those in the reference data (used for deriving 
performance measures). For example, if performance 
measures are derived from reference data set in a rural 
landscape, then in practice those measures cannot be 
applied to gauge filter performance in urban areas. 
e Reference data is usually only available for areas that can 
be measured photogrammetrically or by conventional 
survey. For example, in reference data, areas covered by 
dense vegetation are usually not sampled. 
Nonetheless, filters are developed with the expectation that they 
will succeed. This expectation is founded on the knowledge that 
in most cases the characteristics of the terrain (e.g., slope, 
roughness, form, curvature, etc.,) are bounded and that surfaces 
that fall outside these bounds are not terrain. Figure 1 depicts the 
current approach to filtering. From a monotonic function (based 
on a test point and its neighborhood) a decision measure is 
derived. Based on a pre-selected threshold for the decision 
measure the test point is classified as either terrain or object. As 
already stated this strategy works in most cases. However, it 
also leads to some misclassifications. Therefore, improving 
filtering strategy requires that the reason for misclassification be 
known. Two types of misclassification are possible, Type I 
errors and Type II Type I errors being the 
misclassification of terrain points as object points and Type II 
errors being the misclassification of object points as terrain 
points. This is also shown in Figure 1. Within a certain band, 
either side of the threshold a point has a likelihood of having 
been misclassified (hence, the question marks in Figure 1). 
Knowing where and what type of misclassification has occurred 
can be determined using reference data. However, in normal 
practice this reference data will not exist. Therefore, the 
challenge is to detect misclassification (or the likelihood of it) 
where no reference data exists, or to be more precise, devise an 
alternative means checking the possibility of misclassification. 
errors. 
When reference data does not exist, an operator has to manually 
check the results of the filtering. However, an operator using 
just the point-cloud cannot achieve a 100% classification. There 
A - 330