DISTANTEetal:ACTIVERANGEIMAGINGDATASET... AnnalsoftheBMVAVol.2010,No.

3,pp116(2010)

Active Range Imaging Dataset for Indoor Surveillance

Cosimo Distante1, Giovanni Diraco2, Alessandro Leone2 1 National Institute for Applied Optics INOA Viale della Libert 3 Arnesano, 73010 Lecce (Italy) 2 Institute for Microelectronics and Microsystems National Research Council Provinciale per Arnesano, Palazzina A3, 73100 Lecce (Italy) <[email protected]>

Abstract Range Imaging (RIM) is a new suitable choice for measurement and modelling in many different applications. The ability to describe scenes in three dimensions opens new scenarios, providing new opportunities in different fields, including visual monitoring and security. Active range sensorsprovidedepthinformationallowingtodevelopalgorithmsmuchless complexandallowsproblemstobeapproachedinanew,robustandefficient way. The paper presents a wide dataset for indoor surveillance applications acquiredbyastateoftheartrangecamera.Themainissueisthedefinitionof acommonbasisforthecomparativeevaluationoftheperformanceofvision algorithms.

1 Introduction
Acquiring 3D information in real environments is an important task for many computer visionapplications(i.e.surveillance,security,autonomousnavigation,computergraphics, humancomputerinteraction,automotive,biomedical,augmentedenvironments,etc.)and benefitsfromrangeimageacquisitionareapparent.Inthelastyearsseveralactiverange sensors having small dimensions, lowpower consumption and realtime performances have been released. Since the TimeOfFlight (TOF) range cameras are widely used by researchersinmonitoringandsurveillanceapplications(asdescribedinthenextsection),
2010.Thecopyrightofthisdocumentresideswithitsauthors.
Itmaybedistributedunchangedfreelyinprintorelectronicforms.

DISTANTEetal:ACTIVERANGEIMAGINGDATASET... AnnalsoftheBMVAVol.2010,No.3,pp116(2010)

adhoc workshops on this topic have been organized [1, 2] and emerging projects have beenstarted[3,4,40]inordertodevelopTOFsensorsandrelativesoftwaretechnologies for ambient intelligence (people tracking, head tracking, eye movements tracking, behaviouranalysis,etc.). Comparing algorithms is especially difficult since they are tested on different datasets under widely varying conditions. A large amount of real and synthetic image datasets have been defined for the evaluation of the performance of passive visionbased algorithms in surveillance applications [6, 7, 8, 9], whereas datasets of range images for surveillance and monitoring applications are limited to head/face detection and gesture analysis[15](differentdatabasesarepresentedinliteraturewiththeaimtoevaluateobject segmentationonsyntheticstillrangeimagesandalsoforbiometricsandrobotnavigation [16, 17, 18, 41]). Rather, 3D range imaging datasets for people posture analysis and behaviour recognition are not available yet. For the previous reason, the definition of a database ofreal 3D range images for human behaviour analysis and posture recognition appears useful and important for the scientific community in order to define a common basisfortheevaluationofalgorithmsintheindoorsurveillancefield. This paper presents a large database called ARIDIS (Active Range Imaging Dataset for Indoor Surveillance) of benchmark sequences captured by a stateoftheart TOF range camera.Thepaperisorganizedasfollows.Insection2,thestateoftheartofTOFsensors is discussed and the widely used commercial TOF cameras and related works are presented.Section3providessomedetailsaboutadvantagesanddrawbacksbyusing3D activerangevisioninsteadofpassivestereovisioninsurveillance.Insection4,theARIDIS datasetisdescribedwhereasconclusiveconsiderationsarepresentedinsection5.

2 TOFcameras:StateofTheArtandRelatedWorks
2.1 TOFStateOfTheArt
Optical distance measurement techniques are usually classified into three fundamental categories: Interferometry, Stereo/Triangulation and TimeofFlight [14]. The Interferometrymethodmakesuseoftheperiodicnatureofthelightbyimplementingan interferometricsetupwherethelightfromareferencebeampathinterfereswiththelight reflectedfromanobject.Thehighaccuraciesmainlydependonthecoherencelengthofthe lightsource,sothatInterferometryisnotsuitableforrangesgreaterthanfewcentimetres sincethemethodisbasedontheevaluationofveryshort opticalwavelength.Theactive triangulationmethodexploitsgeometricalrelationsbetweenanilluminationsourceanda camera, whereas the passive triangulation concerns relations between two or more cameras. The geometrical relations between the object, the sensor and a known baseline areusedforthecalculationofthedistance.Inthepassivecasetheacquisitionofdepthis

DISTANTEetal:ACTIVERANGEIMAGINGDATASET... AnnalsoftheBMVAVol.2010,No.3,pp116(2010)

donebystereoscopyor multiviewgeometry,meaningthe clever combination ofat least two 2D images from different points of view. The TimeofFlight technique exploits the finitespeedoflightforthedeterminationofdistancesthroughmodulatedlightsignal.The distribution of distances can be determined by measuring the roundtrip time of the modulatedlightfromtheilluminatortotheobjectandback.Theroundtriptimeofthe

(a)

(b)(c)(d)

Figure 1. a) TOF PMD Principle; b) PMD 19k; c) SwissRanger 3000; d) Canesta PD200.
emittedsignalisrelatedtotheobjectdistanceanditmaybedeterminedinvariousways. Currently, devices employing nanosecond laser pulses are widely investigated; architecturalconstraintsandtheintegrationofhighlypreciseminiaturizedtimingcircuitry increasethecostsofthesystems. The sensors based on the Photonic Mixer Devices (PMD) technology assure advantages according to size, weight and scan time. A PMD sensor employees incoherent, near infrared, amplitudemodulated, continuous wave light and it determinethe signal phase shift(andhencetheobjectdistance)bymixingtheemittedsignalwiththereturnedsignal at every pixel (Fig. 1.a) [13]. The illumination unit of PMD cameras are usually realized withlowcostLEDarrays.VariousTOFcamerassupporttheSuppressionofBackground Intensity(SBI)modalityoranequivalentIRsuppressionscheme,allowingtheusageofthe devicesinoutdoorapplications.Asmentionedabove,theTOFtechniqueisbasedonthe indirect estimation of the arrival time by measuring the phase shift between the

DISTANTEetal:ACTIVERANGEIMAGINGDATASET... AnnalsoftheBMVAVol.2010,No.3,pp116(2010)

transmitted and the received signal. The phase shift is directly proportional to the roundtripdistancedasshowninthefollowingrelation:

d=

mod 2 2

with d

mod
2

(1)

where mod is the modulation wavelength which is about 15m in most TOF cameras. By substituting mod=15m in Eq. 1, the resulting nonambiguous distance is of 7.5m. Recent camerasallowtousemorethanonemodulationfrequenciesinordertoincreasethenon ambiguity range, but increasing device complexity and decreasing accuracy of measure. However,TOFsensorsnormallysufferforthefollowingdrawbacks[29]: LowResolution:Currentsensorshavearesolutionbetween64x48and176x144,which is lower if compared to standard RGB sensors but sufficient for most current applications. Additionally, the resulting large solid angle yields affect the measure in inhomogeneousdepthareas(e.g.atsilhouetteboundaries). DepthDistortion:Sinceasinusoidalsignalispracticallynotachievable,themeasureis affectedbysystematicerrors. Motion Artefacts: Since four phase images are acquired subsequently in order to reconstruct the received signal, movements of the device or in the scene lead to erroneousdistancevaluesatobjectboundaries. MultipleCameras:Theilluminatorofacameracanaffectanothercameraplacedinthe sameenvironment.RecentdevelopmentsallowtheuseofmultipleTOFcameras. Moretechnicaldetailsonrangeimagingcanbefoundin[35].

2.2 CommercialTOFcamera
TheTOFphasemeasurementprincipleisusedbyseveralmanufacturersofTOFcameras, suchasCanestaInc.[30],PMDTechnologiesGmbH[31]andMESAImagingAG[32](Fig. 1.b,Fig.1.c,Fig.1.d).Canestaprovidesseveralmodelsofdepthvisionsensorsdifferingfor pixelresolution,covereddistance,framerateandfieldofview.Canestadistributessensors withfieldofviewranging between30degsand114degs, dependingonthenature ofthe application. Currently, the maximum resolution of Canesta sensor is 160x120. Although someapplicationswouldbenefitfrommorepixels,thesensorcostgrowsinsizewiththe amountofpixels.Canestasensorsaremainlyemployedinautomotiveapplications[10,11] and customized for specific application requirements. The TOF cameras by PMDTechnologiesGmbHincludeallkeytechnologiesforaTOFSystem.Theillumination LEDs are generally mounted in two arrays, one for each side of the camera. This illumination configuration is not optimal [34], since it introduces errors due to TOF differences between the left and the right array that are corrected in hardware. The correlationinsidethecameraisperformedbyusingaCMOSbasedopticalsemiconductor increasing speed and decreasing cost and noise of the system. PMDTechnologies GmbH

DISTANTEetal:ACTIVERANGEIMAGINGDATASET... AnnalsoftheBMVAVol.2010,No.3,pp116(2010)

Specifications
MESA SR3000 176x144 (QCIF) 47.539.6 USB2.0 MESA SR4000 176x144 (QCIF) 43.634.6 PMD Technologies PMD19k 160x120 (QQVGA) 4040 Firewire, Ethernet 3Watt@870nm 918 20,default 7.5 3%ofrange 6cm(std.dev.) at2mand90% reflectivity 15fps No PMD Technologies CamCube 204x204 3232 USB2.0 3Watt@870nm 12 2090 (20,default) 0.37 2%ofrange 3cm(std.dev.) at2mand90% reflectivity 25fps Yes Canesta DPxxx 64x64/120x160 (3088) (30114) USB1.1/2.0 1Watt@785nm 5Watt@870nm 5 13,26,52,104 0.320 2%ofrange 0.6cm20cm (std.dev.)at 0.3m20m 30fps Yes

PixelArraySize FieldofView (HxV) Interface

Illumination Power PowerSupply(V) 12 Modulation 20,default Frequency(MHz) Nonambiguous range(meters) Distance Resolution Distance Accuracy (singlepixel) FrameRate Outdoor 7.5

USB2.0, Ethernet 1Watt@850nm 1Watt@850nm 12 2931 (30,default) 0.35 1%ofrange +/1cm (absolute accuracy) 54fpsmax No

1%ofrange +/1cm (absolute accuracy) 25fps,typical Yes

Table 1. Main features of the TOF cameras employed by scientific community.

provides several models of TOF camera with different features and suitable for measurementsinbrightdaylightsinceallcamerasareequippedwiththeSBItechnology. ThePMDTechnologiescamerascanbeusedtogettheexcellentdepthinformationaswell as gray scale value of the scene. Currently, the PMD devices provide the resolutions of 48x64 (model 13k), 64x16 (model A2) and a nonambiguity distance up to 40m. Greater pixel resolution (160x120) can be achieved by PMD 19k and by the CamCube model (204204)withanonambiguityrangeupto7.5m.However,atnowtheonlycamera modelsreferencedinscientificworksarethe13k,theA2andthe19kmodel,employedin automotiveandroboticsapplications[12,19,20,21]. TheMESAImagingAGhasdevelopedtheSwissRangerTOFcameraseries:SR2,SR3000 and SR4000. MESA SwissRanger cameras are employed in automotive, medical, biometrics,roboticsandsurveillanceapplications[2227].SinceSwissrangerSR3000offers anexcellenttradeoffbetweenpixelresolutionandcovereddistance,ithasbeenemployed for the ARIDIS dataset acquisition. Table 1 summarizes the main characteristics of the discussed optical devices. Since the MESA SR3000 has been employed in dataset acquisition, a more detailed description is discussed in the following [5]. SR3000 camera provides two kind of images at QCIF spatial resolution (a depth map and a grey levels image) at least 15 fps (up to 25fps, variable with setting parameters). The camera works

DISTANTEetal:ACTIVERANGEIMAGINGDATASET... AnnalsoftheBMVAVol.2010,No.3,pp116(2010)

withanintegrated,modulatednearinfraredlightsource(theilluminatoriscomposedby anarrayof55nearinfraredLEDs).Theemittedlightisreflectedbytheobjectsinthescene andsensedbyapixelarrayrealizedinaspecializedmixedCCDandCMOSprocess:the target depth is estimated by measuring the phase shift of the signal roundtrip from the devicetothetargetandback.Moreover,attentionmustbededicatedtotheFieldOfView (FOV) of the device: normally active range sensors exhibit narrow FOV so a pantilt solutioncouldbeusedformonitoringlargeareas.Anarrowbandinfraredfilterisusedso

(a) (b)

(c)

(d) (e)

(f)

Figure 2. Unwanted effects are shown for different acquisition configuration. Noise corrupts the measure (a) since the Integration Time (IT) is too low (IT=5ms). Saturation effects (b) are due to the high IT setting (IT=30ms) and they are limited (c) when the IT is correctly adjusted (IT=20ms). The effects of the IT tuning are visible in the acquired dataset (in f the intensity image). A 10ms IT causes noise in the measure (d), whereas a better tuning (IT=20ms) is shown in (e).
that depth map and intensity image are not affected by environmental illumination conditions (the camera is suitable in night vision applications). By using the default parameters, the SR camera is able to define a depth map with a good approximation (greater than 99%) when the target object falls in the nonambiguity range (up to 7.5 meters). When the cameratarget distance overcomes the nonambiguity range, aliasing effects appear (i.e. object at 8 meters is seenas at 0.5 meters) as shown in Fig. 2.d and Fig. 2.e, demoting only the depth estimation. Another important parameter is the Integration Time (IT), that can be tuned in a proper way to limit saturation and noise effects.Inparticular,iftheITisshort(Fig.2.a)theresultsareverynoisy;ifitislong(Fig. 2.b) the results are smoothed and overflow starts to contaminate the results with objects closetothecamera.TheabovediscussionisalsosuitablefortheotherTOFsensorsusing

DISTANTEetal:ACTIVERANGEIMAGINGDATASET... AnnalsoftheBMVAVol.2010,No.3,pp116(2010)

similartechnologies,astheyexhibitthesameproblems(nonambiguityrange,etc.).A3D plottingofthesceneshowninFig.2.eisprovidedinFig.3.

3 ActiveRangeVisionvs.PassiveVision
Inthissectionwefocustheattentiononadvantagesanddrawbackswhenanactivevision system is used instead of a classical 2D/3D vision system (passive vision). Table 2 synthesizes the most important characteristics of TOF cameras in comparison with triangulationsystemones(passivestereovision).ThemainadvantageintheuseofRIMis the description of a scene with a more detailed information, since RIM provides both depth map and intensity image, that could be used at the same time in the image processing applications. In particular, typical problems of passive vision (foreground camouflage, shadows, partial occlusions, etc.) could be overcome by using depth informationthatisnotaffectedbyilluminationconditionsandobjectsappearance.Onthe other hand, in several situations active RIM exhibits unwanted behaviours due to limitations of specific 3D sensing technology (limited depth range due to aliasing, multi path, reflection object properties). Note that the depth information could be obtained by using inexpensive passive stereo vision. However, this approach presents high computational costs and it fails when the scene is poorly textured or the illumination is insufficient;viceversa,activevisionprovidesdepthmapsevenifappearanceinformation is untextured and in all illumination conditions [28]. It is important to note that both distanceandamplitudeimagesdeliveredbytheTOFcamerahaveanumberofsystematic drawbacks that must be compensated. The main amplituderelated problem comes from thefactthatthepowerofawavedecreaseswiththesquareofthedistanceitcovers.For thepreviousconsideration,thelightreflectedbyimagedobjectsrapidlydecreaseswiththe distance between object and camera. In other words, objects with the same reflectance locatedatdifferentdistancesfromthecamerawillappearwithdifferentamplitudesinthe image.BenefitsofTOFRIMinsurveillancecontextsaresummarizedinTable3,whereas drawbacks are emphasizes in Table 4. Drawbacks are showed in Fig. 4 by using typical framesoftheARIDISdatasetsequences.

DISTANTEetal:ACTIVERANGEIMAGINGDATASET... AnnalsoftheBMVAVol.2010,No.3,pp116(2010)

Depth resolution Spatial resolution Portability Computational efforts Cost

TOFsensor Subcentimetre (if chromaticity conditionsaresatisfied). Medium(aboutQCIF,144x176pixels). Dimensions are the same of a normal camera. Onboard FPGA for phase and intensitymeasurement. High for a customizable prototype (15005000).

Stereo(passive)vision Submillimetre(ifimagesarehighlytextured).

High(over4CIF,704x576pixels). Twovideocamerasareneededandalsoexternal lightsource. High workload (the calibration step and the correspondencessearchprocessarehard). Itdependsonthequalityofstereovisionsystem.

Table 2. Comparison of important characteristics of TOF cameras and stereo vision systems.

Illumination conditions Shadows presence Partial occlusions

TOFsensor Accurate depth measurement in all illuminationconditions. It does not affect principal steps of monitoringapplications. Partiallyoccludedobjectsaredetected as separated (if they are at different depths). Camouflage is avoided but appearance could affect depth precision(chromaticitydependence).

Passivevision Sensible to illumination variations and artificial lights.Unabletooperateindarkenvironments. Reduced performances recognition,etc. in segmentation,

Due to projective ambiguity, occlusions are difficult to detect (merging blobs) and handle (predictivetrackingstrategies). Camouflage effects are presented foreground/background present appearanceproperties. when same

Objects appearance

Table 3. Advantages in the use of TOF sensors in surveillance contexts.

Drawbackdescription Itaffectsthenonambiguityrangei.e.themaximumachieveddepthisreduced(upto7.5 meters). Depthmeasurementisstronglycorruptedwhenthetargetsurfacepresentscorners. Materials having different colors exhibit dissimilar reflection properties that affect reflectedlightintensityand,therefore,depthresolution. Usually it is limited so that an accurate positioning of the sensor is needed. A pantilt architecturecouldbeuseful.

Aliasing

Multipatheffects Objectsreflection properties Fieldofview

DISTANTEetal:ACTIVERANGEIMAGINGDATASET... AnnalsoftheBMVAVol.2010,No.3,pp116(2010)

Table 4. Drawbacks in the use of TOF sensors in surveillance contexts.

(a)(b)

Figure 4. The aliasing effect (a) appears in regions for which the distance from the camera overcomes the 7.5m (the opened door). The multi-path effects are apparent (a) in the regions in which adjacent walls intersect. Moreover, depth measurement is corrupted by object reflection properties (the LCD TV in (a)). Note that the corresponding intensity image (b) doesnt presents the previous discussed problems.
In order to understand the advantages in the use of range imaging in surveillance, a qualitative comparison between intensitybased and depthbased segmentation is presented in Fig. 5, when the same wellknown segmentation approach (Mixture of GaussiansbasedbackgroundmodellingandBayesianframeworkforsegmentation)[39]is used. The better segmentation is achieved by using the depth image, whereas the same segmentationapproachappliedontheintensityimagesuffersofmimeticeffects.

(a)

(b)

(c) (d) Figure 5. Segmentation results are shown (b, d) when the segmentation approach in [39] is applied on depth (a) and intensity (c) information, respectively. The better segmentation is achieved starting from the depth image, whereas the same segmentation approach applied on the intensity image suffers of mimetic effects (sweater and wall present the same brightness).

10

DISTANTEetal:ACTIVERANGEIMAGINGDATASET... AnnalsoftheBMVAVol.2010,No.3,pp116(2010)

4 DatasetsDescription
In the visualsurveillance context, several benchmark sequences and groundtruth data have been defined by using traditional monocular (passive) vision systems (ETISEO [2], CAVIAR [3], PETS [4]). Datasets for the evaluation of range imaging algorithms on dynamic scenes are notavailable yet. The goal of ARIDIS is to define a common dataset thatcouldbeusedinordertoevaluatetheperformanceofcomputervisionalgorithmsby using active RIM technologies in indoor scenes. The main idea is to investigate RIM approaches when typical troublesome issues of the traditional passive vision occur (illumination conditions changing, shadows, mimetic and untextured regions, occlusions asdescribedinTable3).Moreoverthedatasetisusefulfortheevaluationofactivevision basedalgorithmswhentypicalproblemsfortheactiverangeimagingarepresent(aliasing, reflectivesurfaces,multipatheffectsasdescribedinTable4).DuetothelimitedFOVand the nonambiguity range of the device, the benchmark sequences have been captured in smallindoorenvironments(upto55meters). ThecameraconfigurationemployedindatacollectionisshowninFig.6,accordingtothe followingnotation: L:coveredroomlength(5.0meters) H:distancecamera/groundfloor :cameratiltangle :camerarollangle (x,y,z):cameracoordinates The camera is statically positioned in wall mounting configuration at height H from the ground floor with a tilt angle and roll angle . For each collected sequence, a text file provides the parameters (H, , ). The datasetis composed by 61 sequences and each of themisabout1800frameslong,capturedatdifferentframerateandaccordingtothemost critical parameter (Integration Time). In order to cover a large amount of events, several personshavebeencapturedwithdifferentorientationsinthesceneanddifferent

Figure 6. Wall mounting camera setup used in the data collection. The metric information are referred to the TOF camera coordinate axes (x, y, z) represented. The setup parameters (H, , ) are provided for each sequence collected in the ARIDIS dataset.

DISTANTEetal:ACTIVERANGEIMAGINGDATASET... AnnalsoftheBMVAVol.2010,No.3,pp116(2010)

11

Behaviour/Posture Walking/Stand Sequences 1,2,5,6,9,10,11,12,13,14,17,18,19,20,21,22,25,26,29,30,33, 34,35,36,37,38,39,40,43,44,47,48,49,50,51,52,53,54,55,56, 57,58,59,60,61 11,12,13,14,43,44,51,52,53,54,55,56,58,60,61 19,20,59 21,22,33,34,37,38,39,40,47,48,49,50,51,52,53,54,55,56,57 19,20,35,36,37,38,53,54,55 17,18,21,22,25,26,29,30,33,34,39,40,47,48,49,50,55,56,57, 58,59 19,20,35,36,37,38,43,44,51,52,53,54,55,56,57

Falling/Lyingdown Limping Sitting/Sit Crouching/Squat Objectmoving Bending/Bent

Table 5. Behaviours and postures collected in ARIDIS sequences.

heights/dimensions and coloured dresses. Sequences present typical indoor situations, suchaswalkingpeoples(one/morepeopleinthescene),personshavingdifferentpostures (stand, sit, lying down, bent, squat) and persons having different behaviours (walking, fallingdown,limping,sitting,crouching,objectsmoving,bending). Furthermore, the dataset acquisition has been realized in different occlusion conditions: without occlusions, with static occlusions (an object occluding a person, named object person occlusion) and with dynamic occlusions (two or more people in the scene occluding each others, denoted as personperson occlusion) [36, 37, 38]. Tab.5 and Tab.6 report the sequences in ARIDIS dataset grouped by behaviours/postures and occlusion types, respectively. Others useful parameters can be found in the documentation downloadable together with the dataset [33], such as the Integration Time of the device, thecamerasetupparameters,theamountofpersonsinthesceneandthekindofobjects presentinthesceneforeachsequence.Alsoabackgroundmodellingsequenceisprovided whenthebackgroundofthescenechanges.InFig.7keyframesareshowncorresponding to the collected posture/behaviour typologies. The proposed dataset do not includes groundtruthdatareferredtosegmentationandtracking.Furtherworkisaddressedwith the aim to define groundtruth allowing a quantitative performance comparison of algorithms.

Occlusions None ObjectPerson PersonPerson

Sequences 1,2,11,12,19,20,60,61 5,6,17,18,21,22,25,26,29,30,33,34,35,36,37,38,39,40,43,44,47,48, 49,50,51,52,53,54,55,56,57,58,59 9,10,13,14,43,44

Table 6. ARIDIS sequences grouped by occlusion (if any) type.

12

DISTANTEetal:ACTIVERANGEIMAGINGDATASET... AnnalsoftheBMVAVol.2010,No.3,pp116(2010)

(a)

(b)

(c)

(d)

(e)

(f)

(g)

(h)

(i)

(j)

(k)

(l)

Figure 7. Keyframes of the collected postures and behaviours (depth and intensity images). a) Objectperson occlusion; b) Person-person occlusion; c) Walking; d) Falling down; e) Sitting; f) Crouching; g) Object moving; h) Bending. Postures and behaviours are collected in scenes with chairs (i), a desk table, a chair and a stool (j), a large table, chairs and a stool (k) and a vertical desk (l).

DISTANTEetal:ACTIVERANGEIMAGINGDATASET... AnnalsoftheBMVAVol.2010,No.3,pp116(2010)

13

5 Conclusions
Inindoorsurveillanceapplications,rangeimagesprovideabetterperceptionofscenesin all illumination conditions, deterring the use of cheap stereo systems that fail in dark or lowtextured environments. If critical parameters of TOF sensor are adjusted, reliable, computationally lowcost and realtime segmentation/tracking can be realized by only using depth measure, since intensity images present unwanted fluctuations. Depth information overcomes projective ambiguity, whereas intensity image provides appearanceinformation,sothatthejoineduseofthemcouldimprovecriticalsteps(object recognition,behaviouranalysis,etc.)allowingabetterdescriptionofmovingobjects.The suggested dataset provides a common basis to investigate vision algorithms; it can be improved by defining groundtruth data to quantify performances. Further work is addressed with the aim to define groundtruth allowing quantitative performance analysis.

References
[1] Workshop on Time of Flight Camera based Computer Vision (TOFCV), IEEE International Conference on Computer Vision (CVPR 2008), Anchorage, Alaska (USA), June2008. [2]WorkshoponDynamic3DImaging,31AnnualSymposiumoftheGermanAssociation forPatternRecognition(DAGM2009),Jena(Germany),September2009. [3]ARTTSUEProject,http://www.artts.eu/Lastaccessed2November,2009 [4]NETCARITYUEProject,http://www.netcarity.org/Lastaccessed2November,2009 [5]T.Oggier,M.Lehmann,R.Kaufmannn,M.Schweizer,M.Richter,P.Metzler,G.Lang, F.Lustenberger,N.Blanc,Anallsolidstateopticalrangecamerafor3Drealtimeimaging with subcentimeter depthresolution (SwissRanger), Proc. SPIE Vol. 5249, pp. 634545, 2003. [6] A.T. Nghiem, F. Bremond, M. Thonnat, V. Valentin, ETISEO, performance evaluation forvideosurveillancesystems,Proc.IEEEAVSS2007,pp.576481,2007. [7] The CAVIAR website http://homepages.inf.ed.ac.uk/rbf/CAVIAR/ Last accessed 2 November,2009. [8]http://pets2006.netLastaccessed2November,2009. [9]http://marathon.csee.usf.edu/range/rawimagedataLastaccessed2November,2009. [10]O.Gallo,R.Manduchi,A.Rafii,RobustCurbandRampDetectionforSafeParking UsingtheCanestaTOFCamera,IEEEComputerSocietyConferenceonComputerVision andPatternRecognitionWorkshops(CVPRW08),Anchorage(Alaska),June2008. [11] S. Acharya, C. Tracey, A. Rafii, System design of timeofflight range camera for car park assist and backup application, IEEE Computer Society Conference on Computer VisionandPatternRecognitionWorkshops(CVPRW08),Anchorage(Alaska),June2008.

14

DISTANTEetal:ACTIVERANGEIMAGINGDATASET... AnnalsoftheBMVAVol.2010,No.3,pp116(2010)

[12]T.Ringbeck,B.Hagebeuker,A3DTimeofflightcameraforobjectdetection,Optical 3DMeasurementTechniques,ETHZrich,2007. [13]W.Karel,IntegratedRangeCameraCalibrationUsingImageSequencesfromHand Held Operation, XXIst Congress of the International Society for Photogrammetry and remotesensing(ISPRS2008),Beijing(China),July2008. [14] B. Buttgen, P. Seitz, Robust Optical TimeofFlight Range Imaging Based on Smart Pixel Structures, IEEE Transactions on Circuits and Systems, Vol. 55, Issue 6, pp. 1512 1525,July2008. [15]ARTTS3DTOFDatabase,http://www.artts.eu/publications/3d_tof_dbLastaccessed2 November,2009. [16] A. Birk, K. Pathak, S. Schwertfeger,W. Chonnaparamutt, The IUB Rugbot: an intelligent,ruggedmobilerobotforsearchandrescueoperations,InternationalWorkshop onSafetySecurityandRescueRobotics(SSRR)IEEEPress,2006. [17]A.Birk,S.Markov,I.Delchev,K.Pathak,AutonomousRescueOperationsontheIUB Rugbot International Workshop on Safety Security and Rescue Robotics (SSRR) IEEE Press,2006. [18] Jacobs University, Robotics group, http://robotics.iubremen.de Last accessed 2 November,2009. [19] S. Hussmann, T. Liepert, ThreeDimensional TOF Robot Vision System, IEEE TransactionsonInstrumentationandMeasurement,Acceptedforfuturepublication. [20]C.Joochim,H.Roth,Developmentofa3Dmappingusing2D3Dsensorsformobile robot locomotion, IEEE International Conference on Technologies for Practical Robot Applications(TePRA2008),Nov.2008. [21] T. Ringbeck, B. Hagebeuker, A 3D Time Of Flight Camera For Object Detection, Optical3DMeasurementTechniques,ETHZrich. [22] J.T. Thielemann, G.M. Breivik, A. Berge, Pipeline Landmark Detection for Autonomous Robot Navigation using TimeofFlight Imagery, IEEE Computer Society Conference on Computer Vision and Pattern Recognition Workshops (CVPRW 2008), June2008. [23] S.A. Guomundsson, R. Larsen, H. Aanaes, M. Pardas, J.R. Casas, TOF Imaging in Smart Room Environments towards Improved People Tracking, IEEE Computer Society ConferenceonComputerVisionandPatternRecognitionWorkshops(CVPRW2008),June 2008. [24] M.B. Holte, T.B. Moeslund, P. Fihl, Fusion of Range and Intensity Information for View Invariant Gesture Recognition, IEEE Computer Society Conference on Computer VisionandPatternRecognitionWorkshops(CVPRW2008),June2008. [25]YoudingZhu,B.Dariush,K.Fujimura,Controlledhumanposeestimationfromdepth image streams, IEEE Computer Society Conference on Computer Vision and Pattern RecognitionWorkshops(CVPRW2008),June2008.

DISTANTEetal:ACTIVERANGEIMAGINGDATASET... AnnalsoftheBMVAVol.2010,No.3,pp116(2010)

15

[26] J. Penne , C. Schaller , J. Hornegger, T. Kuwert, Robust realtime 3D respiratory motiondetectionusingtimeofflightcameras,InternationalJournalofComputerAssisted RadiologyandSurgery,pp.427431,July2008. [27] T. Oggier, B. Bttgen, F. Lustenberger, G. Becker, B. Regg, A. Hodac, SwissRanger SR3000 andfirst experiencesbasedonminiaturized3DTOFcameras,Proceedingsof the 1stRangeImagingResearchDay.Zurich,Switzerland,2005. [28]Z.Hu,T.Kawamura,K.Uchimura,APerformanceReviewof3DTOFVisionSystems inComparisontoStereoVisionSystems,StereoVision,InTechEducationandPublishing, November2008. [29] A. Kolb, E. Barth, R. Koch, ToFSensors: New Dimensions for Realism and Interactivity, IEEE Computer Society Conference on Computer Vision and Pattern RecognitionWorkshops(CVPRW2008),June2008. [30]CanestaInc.,www.canesta.comLastaccessed2November,2009. [31]PMDTechnologiesGmbH,www.pmdtec.comLastaccessed2November,2009. [32]MesaImagingAG,www.mesaimaging.chLastaccessed2November,2009. [33] SIPlab Imm Cnr, http://siplab.le.imm.cnr.it/3D/index.htm Last accessed 2 November, 2009. [34] H. Rapp, Experimental and Theoretical Investigation of Correlating TOFCamera Systems, Master Thesis, Interdisciplinary Center for Scientific Computing (IWR), UniversityofHeidelberg,September2007. [35]B.Bttgen,T.Oggier,M.Lehmann,R.Kaufmann,F.Lustenberger,CCD/CMOSlock inpixelforrangeimaging:Challenges,limitationsandstateoftheArt,1stRangeImaging ResearchDays,ETH,Zurich,Switzerland,pp.2132,Sep.2005. [36] C.J. Lin, J.G. Wang, Human Body Posture Classification Using a Neural Fuzzy NetworkbasedonImprovedParticleSwarmOptimization,8thInternationalSymposium onAdvancedIntelligentSystems,SokchoCity,Korea,pp.414419,Sep.58,2007. [37]NooritawatiMdTahir,AiniHussain,SalinaAbdulSamad,HafizahHusain,Posture recognition using correlation filter classifier, Journal of Theoretical and Applied InformationTechnology,2008. [38]B.B.Francois,F.Bremond,M.Thonnat,I.S.Antipolis,HumanPostureRecognitionin Video Sequence, Joint IEEE International Workshop on Visual Surveillance and PerformanceEvaluationofTrackingandSurveillance(VSPETS2003),pp.2329,2003. [39]D.Lee,EffectiveGaussianMixtureLearningforVideoBackgroundSubtraction,IEEE TransactionsonPatternAnalysisandMachineIntelligence,Vol.27,no.5,pp.827832,2005. [40] COGAIN Network of excellence, http://www.cogain.org Last accessed 2 November, 2009. [41]G.Passalis,I.A.Kakadiaris,T.Theoharis,G.Toderici,T.Papaioannou,Towardsfast 3D ear recognition for reallife biometric applications, avss, pp.3944, 2007 IEEE

16

DISTANTEetal:ACTIVERANGEIMAGINGDATASET... AnnalsoftheBMVAVol.2010,No.3,pp116(2010)

Conference on Advanced Video and Signal (http://vislab.ucr.edu)Lastaccessed2November,2009. Based Surveillance, 2007.