Data Vsualzaton wthn Urban Models Anthony Steed 1, Salvatore Spnello 2, Ben Croxford 3, Rchard Mlton 1 1 Department of Computer Scence, Unversty College London, UK 2 LaBRI, Unversté Bordeaux, France 3 Bartlett School of Archtecture, Unversty College London, UK A.Steed@cs.ucl.ac.uk, spnello@labr.fr, B.Croxford@ucl.ac.uk, R.Mlton@cs.ucl.ac.uk Abstract To complement the Ordnance Survey data we are Models of urban envronments have many uses for town plannng, pre-vsualzaton of new buldng work and utlty servce plannng. Many of these models are three-dmensonal, and ncreasngly there s a move towards real-tme presentaton of such large models. In ths paper we present an algorthm for generatng consstent 3D models from a combnaton of data sources, ncludng Ordnance Survey ground plans, aeral photography and laser heght data. Although there have been several demonstratons of automatc generaton of buldng models from 2D vector map data, n ths paper we present a very robust soluton that generates models that are sutable for real-tme presentaton. We then demonstrate a novel polluton vsualzaton that uses these models. 1. Introducton As desktop machnes become faster, the compulson amongst computer graphcs researchers seems to be to generate larger models that can brng the new machnes to a grndng halt. Models of the urban envronment are easy test cases as t s relatvely smple to generate very complex models from readly avalable 2D map data. In ths paper we descrbe a process for automatcally generatng 3D models from 2D map data, aeral photography and LIDAR nformaton, and then ntegrate polluton data vsualzatons wthn those models. Urban models can be generated from a varety of dfferent data sources, a survey can be found n [9]. Many of the methods descrbed n that survey assume that the surveyor or modeler starts wth no data and must scan and capture the complete model that they requre. However, n the UK, Ordnance Survey produce extremely good 2D vector data for the whole country. Ths data s kept up to date by teams of surveyors. It s mantaned from Photogrammetrc and surveyng processes. Fgure 1: A representaton of the stages of urban modelng usng LIDAR (LIght Detecton And Rangng) data and aeral photography. LIDAR gves spot heghts at reasonably dense spacng, so as to gve a terran heght and buldng heghts. Both, LIDAR and Ordnance Survey data are vector data. They are really sutable for storage and 2D vsualzaton but they do not have an explct 3D structure. Therefore, we have chosen to make a Constraned Delaunay Trangulaton, bult on those nput data sets, to obtan robust structured nformaton. The resultng 3D models have been desgned to be run wthn real-tme renderers. Thus we have pad attenton to optmzng the number of polygons. The resultng models have been used n a number of applcatons wthn the Equator Cty project [6] where we have been usng urban envronments for 3D vrtual tour gudes. In ths paper we wll descrbe an extenson of the orgnal model for vsualzng urban polluton. Ths s part of Advanced Grd Interfaces for Envronmental e- Scence, whch s assocated wth the Equator IRC [5] [7].
tq3181se tq3180ne 2. Data Modellng tq3281sw tq3280nw 500 m Fgure 2 A representaton of the orgnal NTF data 2.1. System Overvew and Requrements For real-tme modellng we have some strngent requrements on the types of model. The prmary requrement s that the models respect the geometry of the orgnal data but have a good vsual appearance. It s desrable for lghtng and shadow projecton that the buldng foot-prntouts are cut out of the ground plane so that there are no T-junctons on the ground plane. Ths does mean that a larger number of polygons are requred, but t does remove a large number of vsual artefacts. We also requre that any back-facng or non-vsble polygons are removed from the resultng models 2.2. Parttonng the Ground Plane We used three prncpal resources: Ordnance Survey (OS) vector maps, LIDAR (LIght Detecton And N 500 m E Rangng) heght data, and aeral photography. In addton we have started to ntegrate procedures for modellng facades of buldng. Fgure 2 shows four example vector maps from the Land-Lne data set. All maps are Crown Copyrght. Land-Lne s suppled as tled data, wth each tle comprsng 500m x 500m. These are dstrbuted n Natonal Transfer Format (NTF), and we can use ether ths or the OpenGIS Consortum s Geography Markup Language (GML). These maps contan the topography as vector data for both tangble and ntangble features. An ntangble feature s a map feature that does not represent a realworld object e.g. the lne representng a county boundary. The maps contan pont data to represent Spot Heghts, Trangulaton Ponts etc and lne data to represent Buldng Outlnes, Publc Road Central Lne etc. There s no area type, so areas such as buldngs are defned usng lnes wth a unque seed pont to dentfy the area. Feature postons are measured n Natonal Grd coordnates. Features wll have a category code denotng ther type (Spot Heght, Buldng Outlne etc). Features may have an assocated text strng to ndcate the Road Name, House Name, etc. To satsfy our requrement to unquely classfy the ground plane and to remove ambgutes, we frst buld a complete Constraned Delaunay Trangulaton from the vertces. Delaunay Trangulatons are partcular trangulatons, bult on the nput data set, whch satsfes the empty crcum-crcle property: the crcum-crcle of each trangle n the trangulaton does not contan any nput ponts. They take n the gven nput data set and return a structure descrbng the data set. Even f OS vector maps sometmes contan errors or ambgutes such as mssng edges, usng the trangulaton and some feature detals n the orgnal vector map lke buldngs seeds ponts and roads central lnes, we can easly classfy the resultng trangles and edges nto varous sets. For our applcaton we group these n to: Buldngs, Pavements, Water and Roads (see Fgure 3).
Fgure 4 Smooth surface created by nterpolatng the LIDAR data by the nverse dstance weghtng functon Fgure 3 Constraned Delaunay Trangulaton 2.3. Terran Heght At ths stage, the model comprses a planar floor. It s sutable for extruson nto a 3D model. Ths s easly done at ths stage, because n the Delaunay Trangulaton each vector lne has been unquely attached to two planar trangles, so edges of buldngs can be duplcated, one edge rased and the façade polygons nserted. However, f we can obtan heght data we have to consder t before extruson. The Land-Lne data contans spot heghts, but these are too sparse to construct a smooth surface. A better data source, f t s avalable, s LIDAR. Ths gves spot heghts at denstes of typcally 30 ponts per 100 square meters (50 tmes hgher than Land-Lne). The horzontal accuracy s 1.5m n the worst case due to uncertantes related to the atttude of the survey arcraft. The vertcal accuracy s about +/- 15cm. Ths data can be used to gve both the heght of buldngs n the prevous extruson step and to construct a terran heght for the ground. Dfferent methods could be used to create a smooth terran surface from an unorgansed scattered set of data ponts. The most used technques are: Krgng and Inverse Dstance Weghtng (IDW) nterpolaton. Krgng s a method of nterpolaton, whch predcts unknown values from data observed at known locatons. Ths method uses varogram to express the spatal varaton, and t mnmzes the error of predcted values, whch are estmated by spatal dstrbuton of the predcted values [13]. Krgng s a powerful method; unfortunately the calculatons necessary to perform t have a hgh Fgure 5 Trangles comprsng a buldng façade computatonal complexty. Results obtaned wth the second method (IDW) are comparable wth those obtaned wth Krgng n addton IDW mplementaton s easer and ths algorthm s faster than Krgng. For these reasons we decded to use IDW to fnd the heght of each pont on the ground plane. IDW s descrbed n detals n Secton 3.4. See Fgure 4 for an example result. 2.4. Extruson Secton 2.2 ntroduced a procedure to fnd buldngs outlne even n maps contanng errors or ambgutes. The basc extruson algorthm descrbed n ths paragraph takes a buldng outlne and extrudes ts edges. The extruson s done n three parts. Two parts are composed by a ground secton, whch s one storey hgh,
and an upper secton, whch flls from the 1 st storey to the roof level. These are the pnk and green strps n Fgure 5. Ths s done so that a texture map wth doors and wndows can be used for the ground storey, and wth wndows only for the upper floors. The extracton procedure must consder the terran model (see precedent Secton) to make more realstc buldngs. In effect, the terran surface normally s not planar, then, ndvdual buldngs may not have horzontal edges wth the ground. Therefore, a thn polygonal strp s nserted around the base of the buldng as shown Fgure 5 (blue strp). We have chosen to model Buldngs wth multple levels so that each level can be textured dfferently. At the current tme, we are modellng only flat roofs, so to fnd the heght of a buldng we fnd the hghest LIDAR pont wthn the buldng footprnt, or a close-by but above street-level LIDAR pont f the buldng footprnt s small and there s no LIDAR heght pont wthn t. Water features need specal treatment, snce they wll rarely be planar due to surveyng and nterpolaton processes. Water features are flattened by fndng ther lowest feature pont. The water feature s then flattened to ths heght and approprate walls put n where the water feature adjons other features. 2.5. Other Model Features If aeral photography exsts, then t can be draped over the mesh. We have used sectons of the Ctes Revealed data set from GeoInformaton Internatonal for current demonstratons. Snce the pxel sze of typcal aeral photography s around 1 pxel/m, there wll be obvous bleedng of ground features to roofs and vce versa. The Land-Lne data contans nformaton about pont features such as street furnture and trees. These can be modelled and nserted. Trees are problematc snce they also appear on the aeral photography. Fnally pre-modelled buldngs can be nserted. Ths nvolves some preparaton. The orgnal 2D vector data has to be marked so that the polygons are not extruded. So far we have not dealt wth fttng the footprnt of the pre-modelled buldng wth 2D map, so we smply leave the buldng outlne as a ground plane. 2.6. Other Outputs Snce we have bult a complete Delaunay trangulaton of the ground plane, we can robustly classfy any new pont nto a ground coverage type (road, pavement, buldng, water, etc.). We can use ths to create consstent btmap representatons of the model, wth, say all buldngs classfed. Ths s useful for real-tme CO (ppm) 7 6 5 4 3 2 1 0 0 20 40 60 80 100 120 Tme (s) Fgure 6 Raw data from a segment of a path near UCL collson detecton of avatars wth models, as demonstrated n the system of Teccha et al. [17]. From the Delaunay trangulaton, we can also construct graphs of road and pavement connectvty by followng the mesh connectons for a partcular ground coverage type. These can be used for path plannng for walkng or drvng smulatons. For example, the pavements defne walkable surfaces for avatars, and the pavement graph can be used to smplfy path searchng. In [12] Loscos et al. dscuss how to augment the pavement graph network wth lkely road crossng nformaton so that avatars can walk across the whole map. 3. Ar Qualty Vsualsaton 3.1. Ar Qualty Informaton The pollutant we are studyng s carbon monoxde. Transport makes the greatest contrbuton to carbon monoxde levels and carbon monoxde affects urban areas more sgnfcantly than rural areas. Overall carbon monoxde levels have fallen snce the 1970s, averagng 1mg/m 3 [4]. The Ar Qualty Ste contans archve data from over 1500 UK montorng statons gong back n some nstances to 1972 [2]. Such data sources gve a good pcture of varaton from urban to rural areas. In urban areas some sense of potental varaton s conveyed by the dfference n readngs between kerbsde sensors and sensors placed n background areas away from pollutant sources. However they don t capture the detal of per street varaton. Carbon monoxde dsperses over a matter of hours, but Croxford et al. have shown that ths s affected by local street confguraton [1]. Ths study used a cluster of sensors n fxed placements n a small area around
Unversty College London (UCL). The Ar Qualty Strategy for England, Scotland, Wales and Northern Ireland [3], suggests a standard of 10ppm (11.6mg/m3) runnng 8-hour mean. In the vcnty of UCL, the Croxford study found a peak CO concentraton of 12ppm, but nearby sensors reported much lower values near the background level for CO. Thus, movng pedestrans or vehcles would probably not experence ths peak for a long perod. 3.2. Mappng Ar Qualty In the Equator IRC e-scence project Advanced Grd Interfaces for Envronmental Scence n the Lab and n the Feld (EPSRC grant GR/R81985/01) [7], we have been nvestgatng ways of mappng polluton usng tracked moble sensors [16]. An accurate carbon monoxde sensor s coupled wth a GPS recever and a loggng devce. Ths devce can be ftted nto a bag or placed on a bke rack. The devce logs tme, poston and polluton level. The resultng recordngs are less accurate, but potentally from a wde area of samplng. Wth several such devces beng carred around, t wll be possble to buld a map that shows detaled local varatons n polluton. 3.3. Raw Polluton Data The data shown n Fgure 6 was collected on a path startng n UCL s front Quad, and walkng up towards Euston Road. Before reachng Euston Road, the user crossed to the other sde of the road, and the peak was reached when they were stood near the traffc lghts at the juncton of Euston Road and Gower St. The peak capture was 6.1 ppm. 3.4. Data Modellng The nput data for the polluton model s a stream made of a GPS poston (x, y ) and polluton data f (CO n parts per mllon). To make a 2D feld representaton, we frst extract a temporal secton of the data. The resultng data set s treated as a set of rregular scatter ponts. One of the most commonly used technques for nterpolaton of scatter ponts s Inverse Dstance Weghted (IDW) nterpolaton. IDW methods are based on the assumpton that the nterpolatng surface should be nfluenced most by the nearby ponts and less by the more dstant ponts. The nterpolatng surface s a weghted average of the scatter ponts and the weght assgned to each scatter pont dmnshes as the dstance from the nterpolaton pont to the scatter pont ncreases. The smplest form of nverse dstance weghted nterpolaton s Shepard's method [14]. The equaton used to fnd the value at poston (x,y) s: F n ( x, y) = w ( x, y) = 1 f where n s the number of scatter ponts n the set, f are the prescrbed functon values at the scatter ponts (e.g. the polluton values), and w are the weght functons assgned to each scatter pont. The weght functon s: w ( x, y) h p = n h j j= 1 ( x, y) p ( x, y) where p s a postve real number (typcally, p=2) and ( x y) ( x x ) 2 + ( y y ) 2 h, = s the dstance from the scatter pont.
nterpolates each scatter pont and s nfluenced most strongly between scatter ponts by the ponts closest to the pont beng nterpolated. 4. Results Fgure 7 Overvew of area around St Paul's Fgure 8 Oblque vew of St Paul's area 4.1. Cty Models The results of the cty model generaton are shown n Fgures 7 and 8. These show a model of nne sq km around the St Paul s area of central London. The model comprses 1.2M polygons. Un-optmsed ths renders at 3-4 frames a second on a PC wth GeForce4 graphcs accelerator. An ongong theme of research at UCL s nteracton and nteractve renderng of large-scale urban models [15]. In that demonstraton we bult a renderer that adapted to frame-rate changes by alterng clps volumes and level of detal. The models descrbed n ths paper are much superor n detal and geometry to the models used n the prevous demonstrator. 4.2. Combned Data Our am n combnng data s two fold: to support vsualsaton by placng the data n the context of the stuaton where t was gathered, and to support remote collaboraton where one partcpant s usng a vrtual envronment dsplay to collaborate wth a colleague n the feld. Fgure 9 shows vews of the juncton between Gower St and Euston Road. The blue lne represents the recorded path from the GPS recever. The naccuracy of GPS locaton can be noted snce the carrer walked along the centre of the pavements except when crossng Gower St. In the vsualsaton n Fgure 9 we present the polluton nterpolaton by colourng the roads. In order to mantan a hgh frame rate, we only nterpolate the polluton level at each vertex of the road polygons usng the nverse dstance weghted nterpolaton. We then use the bult n Gouraud shadng algorthm of standard Fgure 9 Vews of the juncton of Gower St and Euston Road The effect of the weght functon s that the surface
graphcs drvers to do a smooth nterpolaton. Ths typcally uses a b-lnear nterpolaton. For a more accurate vew, a 2D raster mage can be calculated at some fxed spatal frequency. In Fgure 9 the juncton s obvously the most polluted area. 5. Conclusons and Future Work For the modellng work we are contnung by addng roof structure nformaton. Ths can be done by determnng a roof slopes from LIDAR or by estmatng the roof type and then generatng a roof that fts [10]. We are also workng to ntegrate better façade texturng that fts wth the colour and lghtng nformaton from aeral photography. We tend to use ether automatc façade texturng or draped aeral photography at the moment snce the vsual results are often jarrng f both are used together. We are also workng to ntegrate Photogrammetrc procedures for rapd modellng of specfc buldng facades. A second actvty s on buldng a run-tme that can dsplay larger sectons of the vrtual model. For groundlevel exploraton, we can use a combnaton of occluson cullng and mposter-based renderng. The models we have are very sutable for certan types of occluson cullng, because connectvty between buldngs can be recorded and exported. A thrd actvty s to make a full ntegraton wth the crowd anmaton system of Teccha, Loscos, et al. [12] [17]. For the polluton modellng we have demonstrated the feasblty of makng dense maps of polluton usng moble sensng devces. Ths enables new types of montorng that address local varaton n polluton and also the levels of polluton that are experenced by dfferent users of the urban space. We hope to establsh the polluton montorng nfrastructure as a publc nfrastructure that can be shared or nstantated by other users. Further Detals Ths work s supported by the projects Advanced Grd Interfaces for Envronmental e-scence n the Lab and n the Feld (EPSRC Grant GR/R81985/01) and the EQUATOR Interdscplnary Research Collaboraton (EPSRC Grant GR/N15986/01). For a complete overvew of the envronmental e- scence project, ncludng a companon project on envronmental montorng n the Antarctc, see the EQUATOR webste pages [7]. For example data sets and more detaled specfcatons of the devce see the web page [7]. We plan to make a publc release of the software, and to host an example vsualsaton servce at that address. For further nformaton about the pollutonmontorng project please contact Anthony Steed (A.Steed@cs.ucl.ac.uk). References [1] Croxford, B., Penn, A., Hller, B. (1995) Spatal Dstrbuton of urban polluton: cvlzng urban traffc, Ffth Symposum on Hghway and Urban Polluton, May 22-24, 1995. [2] Department for Envronment, Food and Rural Affars (Defra) The Ar Qualty Archve, http://www.arqualty.co.uk/ (verfed 2003-08-13). [3] Department for Envronment, Food and Rural Affars (Defra) (1999) The Ar Qualty Strategy for England, Scotland, Wales and Northern Ireland, 1999, avalable onlne at http://www.defra.gov.uk/envronment/consult/arqual ty/pdf/arstrat.pdf (verfed 2003-08-13). [4] Envronment Agency, Ar Qualty Carbon Monoxde, http://www.envronmentagency.gov.uk/yourenv/eff/ar/222825/222913/?lang =_e (verfed 2003-08-13). [5] EQUATOR, Advanced Grd Interfaces for Envronmental e-scence: Urban Polluton, http://www.cs.ucl.ac.uk/research/vr/projects/envesc/, (verfed 2003-08-13). [6] EQUATOR, Cty Project, http://www.dcs.gla.ac.uk/scrpts/global/equator/mon. cg/ (verfed 2003-08-13). [7] EQUATOR, Envronmental e-scence Project, http://www.equator.ac.uk/projects/envronmental/nd ex.htm (verfed 2003-08-13). [8] EQUATOR, The Equator UnIversal Plaform, http://www.equator.ac.uk/technology/equp/ndex.ht m (verfed 2003-08-13). [9] Hu, J., You, S., Neumann, U. (2003) Approaches to Large-Scale Urban Modelng, IEEE Computer Graphcs and Applcatons, Nov/Dec,2003 [10] Laycock, R.G. and Day, A.M. (2003). Automatcally generatng roof models from buldng footprnts, In WSCG, 2003 [11] Learan Desgn Ltd, http://www.learan.co.uk (verfed 2003-08-13) [12] Loscos, C., Marchal, D., Meyer, A. (2003) Intutve Crowd Behavour n Dense Urban Envronments usng Local Laws, In Theory and Practce of Computer Graphcs 2003, IEEE Computer Socety Press. [13] Olver, M. A. and Webster, R. Krgng: a method of nterpolaton for geographcal nformaton system,
INT. J. Geographcal Informaton Systems, 1990, VOL. 4, No. 3, 313-332 [14] Shepard, D. (1968) A two-dmensonal nterpolaton functon for rregularly-spaced data, Proc. 23rd Natonal Conference ACM, ACM, 517-524. [15] Steed, A., Frecon, E., Pemberton, D., Smth, G. (1999) The London Travel Demonstrator, Proceedngs of the ACM Symposum on Vrtual Realty Software and Technology, December 20-22nd 1999, pp. 50-57, ACM Press. [16] Steed, A., Spnello, S., Croxford, B., Greenhalgh, C. (2003). e-scence n the Streets: Urban Polluton Montorng, UK e-scence All Hands Meetng, September 2003 [17] Teccha, F., Loscos, C., Chrysanthou, Y. (2002) Vsualzng Crowds n Real-Tme. Computer Graphcs forum, 21(4), December 2002, pages 753-765.