AUTOMATIC EXTRACTION OF BUILDING FEATURES FROM TERRESTRIAL LASER
SCANNING
Shi Pu and George Voslman
International Institute for Geo-information Science and Earth Obrvation (ITC)
spu@itc.nl,voslman@itc.nl
Commission VI
KEY WORDS:3D modeling,terrestrial lar scanning,gmentation,feature recognition
ABSTRACT:
Realistic 3D city models are required for many purpos such as urban planning and safety analysis.The traditional modeling methods are mainly bad on manual or mi-automatic reconstruction from clo-range images,which are rather time-consuming.The lack of fully automatic reconstruction is mainly due to the difficulty to recover building structures from clo-range images.T
he terrestrially scanned lar points usually contain uful information,and it can be a valuable data source for reconstructing 3D city models.This paper prents our approach to automatically extract building features from terrestrial lar scanned data.This is done by first processing the terrestrial lar points with various gmentation algorithms,then retrieving veral important properties (size,position,direction,topology,etc.)from the gments,and finally recognizing potential building features (walls,windows and doors,etc.)with feature constraints,which are bad on the properties of gments.The recognized features will form the basis for an automatic 3D building model reconstruction framework.
1
INTRODUCTION
Realistic 3D city models are required for many purpos such as urban planning and safety analysis.Originally the involvement of citizens in urban planning is generally limited to 2D design plans,which may be difficult to interpret.The availability of 3D models of the current urban environment,as well as new urban objects and their alternatives,would increa this involvement remark-ably.3D city models can also play an important role at curity analysis as well.A realistic env
ironment is esntial for making good curity analysis and training,particularly where physical curity (curity of infrastructure)and social curity (livability,curity feeling)are concerned.Figure 1gives an example of the virtual city model of Helmond,the
Netherlands.
Figure 1:VR Helmond
However,nowadays the automatic construction of realistic city model is still not feasible.The traditional modeling methods are mainly bad on manual reconstruction from 2D GIS data and clo-range images,or mi-automatic reconstruction from clo-range images.The manual approach normally starts with building outlines generated from 2D GIS data.Then the 2D outline is sim-ply elevated with certain height to a 3D rough model.Detailed 3D structures can be created on the rough model with commer-cial 3D modeling packages such as 3DStudio Max.Finally the
3D model is textured by manually lecting certain parts from clo-range images.Due to the huge number of urban objects in a city and variety of shapes,manual creation of a city model is a rather time-consuming and expensive procedure.There are also a few rearches on mi-automatic city model reconstruc-tion from clo-range images (Dick et al.2000,Schindler and Bauer 2003),but the results are not satisfactory.This is mainly due to the difficulty to recover 3D building structures from 2D images.Several studies (Brenner 2000,Maas 2001,V oslman et al.2004)show that lar scanning data can be a valuable data source for the automatic city model reconstructing.Comparing to clo-range imagery,terrestrial lar scanning gives explicit 3D information,which enables the rapid and accurate capture of the geometry of a complex building facade;terrestrial lar scanning also provides high density point clouds,which gives enough raw data from which accurate and detail
ed 3D models can be obtained (Rabbani 2006).The reconstruction process bad on terrestrial lar scanning can be generalized as three steps:feature recogni-tion,where important building features (walls,windows,doors,etc.)are extracted;model fitting,where recognized features are fitted to simple geometric shapes such as polyhedron;model re-construction,where models are combined from fitted geometric shapes and other data sources.
This paper prents our approach to extract building features from terrestrial lar scanned data.Section 2gives an overview of our recognition method.Section 3introduces the planar sur-face growing algorithm for gmentation.Section 4describes the recognition procedure,by first defining the feature constraint categories and then giving specific feature constraints for differ-ent features.Sections 5analyzes the recognition quantity.Some conclusions and future work are given in the last ction.
2
METHOD OVERVIEW
A lar scanning point cloud contains information about building facades with x,y,z coordinates.However,the facade structures are not directly understandable by machines.Humans ca
n easily find building features by comparing a t of characteristics such
as position,color,topology,etc.For example,we know the fea-ture”wall”is usually the biggest plane in building facade,and it is usually vertical;we know the feature”window”is on the wall,and it has a certain area range;we know the feature”roof”is above the wall,and it is never vertical.This human knowledge about buildings can be”modeled”and”taught”to machines,so that features can be recognized from point clouds automatically. The feature recognition procedure starts with gmentation,where a point cloud is categorized into different groups so that the points belonging to the same surface or region are in the same group. Each group(gment)is considered a potential building feature and will be analyzed further in the later stage.Next,veral im-portant properties(size,position,direction,topology,etc.)are re-trieved from the gments.A couple of feature constraints are de-fined for each building feature,bad on human knowledge about buildings.In thefinal step,building features are recognized out of gments by checking each gment’s properties through the feature constraints.
3SEGMENTATION
Segmentation is the process of labeling each measurement in a point cloud,so that the points belong
ing to the same surface or region are given the same label.Building features can be roughly extracted from point cloud after gmentation,becau different features usually belong to different surfaces or regions. Several gmentation algorithms bad on lar point cloud are available,and we adopted the planar surface growing algorithm by(V oslman et al.2004)becau it is more suitable for g-menting planar surfaces.We only include a short explanation about the planar surface growing algorithm here becau of its strong relevance to our feature recognition method.
The planar surface growing algorithm consists of the following steps:
1.Determine a ed surface.A ed surface consists of a group
of nearby points thatfit well to a plane.The algorithm -lects an arbitrary unclassified point and tests if a minimum number of nearby points can befitted to a plane.If this is the ca,the points constitute the ed surface.Otherwi, another arbitrary point is tested.
2.Grow the ed surfaces.The growing can be bad on one
or more the following criteria:
•Proximity of points.Only points within certain dis-
tance to a ed surface can be added to this ed sur-
face.
•Globally planar.For this criterion a plane equation
is determined byfitting a plane through all surface
points in this ed surface.Points can only be added if
the perpendicular distance to the plane is below some
threshold.
Figure2shows the original lar scanned point cloud of a build-ing facade,and Figure3shows the gmentation results of this point cloud.It is clear that most windows,doors,roofs and ex-trusions are gmented successfully,although there are still some significant errors,for example,the wall results in2gments,and some parts of windows are not gmented.
In the ideal situation,each gment reprents a building feature, and each building feature is reprented in a gment.However, this can be hardly achieved
becau:
Figure2:A terrestrial lar scanned building
facade
Figure3:Segmentation result of a building facade
•A lar point cloud always contains irrelevant data,such as the objects behind windows,cars,benche
s besides build-ings,and so on.The gmentation results normally contain the irrelevant gments too.
•Segmentation results are not100%correct.Bad gmenta-tion results in over-gmentation(one feature gmented to veral gments),or under-gmentation(veral features gmented to one gment),or miss gmentation(feature is not gmented).
Several parameters need to be specified for the planar surfaced growing algorithm,such as the number of eds,the surface grow-ing radius,the maximum distance between surfaces,etc.With the same number of eds,larger surface growing radius or larger maximum distance between surface lead to fewer gments,which brings under-gmentation.In the other hand,smaller surface growing radius or smaller maximum distance between surface lead to more gment/over-gmentation.Our experience is that it is usually better to over gment lar point than under g-ment,becau over-gmented parts have some similar proper-ties,for example,they are attached to each other,they belong to the same plane,etc.According to feature constraints and the similar properties,over-gmented parts can always be combined to a complete feature in later pha.But under-gmented parts can hardly be split again.
4FEATURE RECOGNITION
The result from gmentation gives potential building features, but it is still unknown whether a gment reprents a feature, or which kind of feature a gment reprents.Each building feature has its own characteristics,which can be formulated as feature constraints that are understandable by machines,so auto-matic feature recognition becomes possible.
春节为什么要拜年
花朵手工制作4.1Feature constraints
Considering the human knowledge about building features,and supposing all the building features are planar,we summarized a t of feature constraint categories as below:
•Size constraint.Walls,windows and doors can be easily distinguished from other features or noi gments by their sizes,as stated before.
•Position constraint.Certain features appear only in certain positions.For example,windows and doors are always on the walls;roofs are always on the top of walls.•Direction constraint.Walls and roofs can be distinguished by their directions,as walls are usually vertical and roofs are not.
•Topology constraint.Building features have certain topol-ogy relation with other features or grounds.For example, ground always intercts some walls;roofs always interct walls.
•Miscellaneous constraint.Some other information can be also helpful to feature recognition.For example,sometimes gments for windows usually have much lower point den-sity,becau glass reflects fewer lar puls than other parts of a building.But this is just an optional constraint category, as sometimes windows are covered with curtains and reflect more lar puls.
4.2Recognition
We list the feature constraints for7important features(ground, wall,window,roof,door,extrusion,intrusion)in Table1.Con-sidering ground is very helpful to recognize building features,we also include ground as one of the features to be recognized,al-though ground itlf is not a building feature.
Properties of a gment,such as area,direction and topology, can be hardly determined directly,as a gment is just a group of points.However,we could approximate each gment with a convex hull of all the points in this gment,and u the prop-erties of this convex hull instead.In other words,the size of a gment is approximated with its convex hull’s area;position is approximated with its convex hull’s geometry center;direction is simply the normal of its convex hull;topology of gments can be approximated with topology of their convex hulls;and so on.Fig-ure5gives an example of the convex hulls of the roof gments in Figure4.
The implementation of door feature recognition is given below as an example.Recognition of other features have the similar structure and their implementations will not be repeated.
1.Inputs:An array of convex hulls of all gments=hulls;
劳务费是什么estimated minimum door area=min;estimated maximum door area=max;convex hull of ground=ground;convex hull of wall=wall;distance threshold=
d
Figure4:3roof
gments
Figure5:Convex hulls of3roof gments
2.Initialize counter i=0
3.while i<hulls.size do
4.if hulls[i].area3D≤min or hulls[i].area3D≥max then
5.Not a door feature.Continue with the next convex hull.
7.if distance(hulls[i].center,wall)>d then
8.Not a door feature.Continue with the next convex hull.
10.if not hulls[i].isVertical then
11.Not a door feature.Continue with the next convex hull.
13.if hulls[i].Intect(ground)=fal then
14.Not a door feature.Continue with the next convex hull.
16.This is a door feature.Push i to door feature list.
For some features,the recognition is bad on other already rec-ognized features,or,in other words,certain features have higher priority during the recognition process.According to Table1, recognition of window,roof,door,extrusion and intrusion require comparing with wall or ground,hence wall features and ground features should be recognized before processing the other5fea-tures.Recognition of walls requires again checking interction with ground,therefore ground feature has even higher priority than wall feature.Recognitions of extrusion/intrusion are depen-dent on wall or roof,which yield another priority adjustment.So finally the recognition priority for the7features should be: Ground>Wall>Roof=Window=Door>Extrusion>Intrusion
Size
Position Direction Topology Miscellaneous
Ground Segment(s)with large area
Lowest
Wall Segment(s)with large area
Vertical
May interct with the ground
Window Area from min <max
On the wall Vertical Low lar points clouds den-sity
Roof Segment(s)with large area
意见Above wall Not vertical Intercts with a
wall
Door Area from min <max
On the wall
Vertical
Intercts with the ground Extrusion A little bit outside the wall/roof
背疼怎么缓解
Intercts with a wall
Intrusion
A little bit inside the wall/roof
Intercts with a
wall
Table 1:Constraints for 7features
4.2.1Ground Although ground is not a building feature,many building features can be easily recognized by comparing with ground.For example,gments of window and door have simi-lar size,position and direction.However,doors always intercts with ground,while most windows usually
don’t.
A ground feature has a relatively large area,and it is the gment with the lowest position.Figure 6shows the recognized ground from the gmented terrestrial lar scanning point cloud in Fig-ure 3.
Figure 6:Recognized ground feature
4.2.2Wall Wall feature is probably the most important build-ing feature,and the determination of many features also depends on the detected wall features.
A wall feature has
a relatively large area,it is vertical,and it intercts with the ground.Figure 7left shows the recognized walls from the gmented terrestrial lar scanning point cloud in Figure 3.
Note that actually 2gments are recognized as walls,which means the wall is over gmented.In this
ca we can automat-ically merge the 2gments into 1(Figure 7right),according to that they face the same direction and are attached to each other.
Figure 7:Recognized wall feature (left:before merge;right:after merge)
4.2.3Roof A roof feature has a large area,it is above the wall,it is not vertical,and it has interction with wall.Figure 4shows the recognized roofs.
4.2.4Door A door features has certain area range,it is on the wall,it is vertical,and it has interctions with ground level.Figure 8shows the recognized doors.
Figure 8:Recognized door feature
4.2.5Window A window feature has certain area range,it is on the wall,it is vertical,and sometimes the point density is lower than average point density for all gments.Figure 9shows the recognized windows after automatically merging over gmented gments.
Figure 9:Recognized window feature
4.2.6Protrusion A protrusion feature is a little bit outside the walls,and it has interction with walls.Figure 10shows the recognized protrusions.
4.2.7Intrusion An intrusion is a little bit inside the walls,and it has interction with walls.No intrusion is recognized from the building facade in Figure 3.
5
QUALITY ANALYSIS
The recognition method is implemented with C++code,and ex-perimented on a PC with Pentium 43.2G CPU,1GB system memory and the NVidia Quadro FX540video card.The tested
Figure10:Recognized protrusion feature
古栈道point clouds contains238034points,and the overall running time is42conds,including18conds for gmentation,7conds for convex hull creation and17conds for feature recognition. We are satisfied with the processing speed.
Total number Recognized number
Ground11
Wall11
Window117
Roof33
Door55
Protrusion33
Intrusion00
Table2:Recognition quality
Table2gives the recognized numbers of the ven features.Ground, wall,roof,door and protrusion are all very well recognized,while only7window features are correctly recognized out of11.This difference in recognition rate is mainly due to the different g-mentation quality.The planar surface growing algorithm always starts with a ed surface,which is a group of nearby points that fit well to a plane.In the growing pha points are added to the ed surface,if the distance of a point to the plane is below some threshold.The plane parameters are updated after every added point,so the larger the plane the more reliable the plane parame-ters will be.That is why large features such as ground or wall will generate more reliable surfaces.However,windows have smaller size and windows without curtain reflect very few lar puls. The points for windows are sometimes not enough to grow reli-able surfaces,and hence windows are only partially gmented, or not gmented at all.
质感背景6CONCLUSIONS AND FUTURE WORK
In this paper we described our approach to automatically ex-tract building features from a terrestrial lar scanned point cloud. First the point cloud is gmented to veral planar parts accord-ing to the point’s location,direction,and the plane it belongs to;then we formulated a ries of feature constraint categories, which reprents the most significant characteristics that can dis-tinguish one f
eature from another and the noi gments;finally each gment is checked through all feature constraints,to deter-mine which kind of feature its is,or just noi gment.
The future work will be mainly on improvement of recognition rate,automatic determination of parameter values,and recogni-tion of non planar features.All the detected features will then be ud in an automatic process of modeling building facades.
ACKNOWLEDGEMENT
英语励志名言短句This rearch is partially funded by the Dutch BSIK rearch pro-gramme Space for Geo-Information,project Virtual reality for ur-ban planning and safety.The authors would also like to thank the Oranjewoud and Cyclomedia for providing the data.
REFERENCES
Brenner,C.,2000.Towards Fully Automatic Generation of City Models,International Archives of Photogrammetry and Remote Sensing,vol.XXXIII-B3,Amsterdam,2000,pp.85-92.
Dick,A.,Torr,P.,and Cipolla,R.,2000.A model-bad method for building reconstruction,Proceeding11th British Machine Vi-sion Conference,Bristol,UK,pp.372-381.
Maas,H.-G.,2001.The suitability of Airborne Lar Scanner Data for Automatic3D Object Reconstruction,Third Interna-tional Workshop on Automatic Extraction of Man-Made Objects from Aerial and Space Images,Ascona,Switzerland.
Rabbani,T.,2006.Automatic Reconstruction of Industrial In-stallations,Publication on Geodesy of the Netherlands Geodetic Commission,vol.62,Phd thesis,TUDelft,Delft,the Netherlands
Schindler,K.and Bauer,J.,2003.A model-bad method for building reconstruction,Proccedings ICCV Workshop on Higher Knowledge in3D Modelling and Motion Analysis,Nice,France, pp.74-82
Suveg,I.and V oslman,G.,2004.Reconstruction of3D build-ing models from aerial images and maps.ISPRS Journal of Pho-togrammetry and Remote Sensing,vol.58(3-4),pp.202-224.
V oslman,G.,Gorte,B.G.H.,Sithole,G.,Rabbani,T.,2004. Recognising Structure in Lar Scanner Point Clouds,Interna-tional Archives of Photogrammetry,Remote Sensing and Spa-tial Information Sciences,vol.46,part8/W2,Freiburg,Germany, October4-6,pp.33-38.
