SLAM-based 3D outdoor reconstructions from lidar data

SLAM-BASED 3D OUTDOOR RECONSTRUCTIONS FROM LIDAR DATA I an Caminal, Josep R. Casas, San iago Royo Dep . d’ ` Op ica i Op ome ia Dep . de Teo ia del Senyal i Comunicacions Uni e si a Poli ` ecnica de Ca alunya ABSTRACT The use o dep h (RGBD) came as o econs uc la ge ou - doo en i onmen s is no easible due o ligh ing condi ions and low dep h ange. LIDAR senso s can be used ins ead. Mos s a e o he a SLAM me hods a e de o ed o indoo en i onmen s and dep h (RGBD) came as. We ha e adap ed wo SLAM sys ems o wo k wi h LIDAR da a. We ha e com- pa ed he sys ems o LIDAR and RGBD da a by pe o ming quan i a i e e alua ions. Resul s show ha he bes me hod o LIDAR da a is RTAB-Map wi h a clea di e ence. Addi- ionally, RTAB-Map has been used o c ea e 3D econs uc- ions wi h and wi hou pho ome y om a isible colo cam- e a. This p o es he po en ial o LIDAR senso s o he econ- s uc ion o ou doo en i onmen s o imme sion o audio i- sual p oduc ion applica ions. Index Te ms—LIDAR came as, mapping, ime-o - ligh , SLAM, 3D imaging, poin -cloud p ocessing 1. INTRODUCTION Simul aneous localiza ion and mapping (SLAM) is he com- pu a ional p oblem o building a map o an unknown en i on- men while simul aneously keeping ack o an agen ’s loca- ion wi hin i . Mapping allows o localize he senso whe eas a loca ion es ima e is needed o build he map. Some SLAM scena ios ocus on loca ion, such as in au omo i e whe e he map uses o be known be o ehand, while in audio isual and special e ec s he ocus is a he on mapping, i.e. econs uc- ion o he scene en i onmen . LIDAR imaging [1] is a powe - ul measu emen echnique whe e a lase pulse is shone on o an objec and he beam e lec ed back is eco e ed a some solid-s a e de ec o . The ime elapsed is measu ed, allowing o an au oma ed measu emen o he dis ance o he a ge , wi hou any u he calcula ion. The concep is also e e ed o as lada o ime-o - ligh imaging. Popula applica ions in- ol e landing aids, objec ecogni ion o sel -guided ehicles. This pape ocuses on adap ing wo s a e o he a SLAM s a egies o wo k wi h LIDAR senso s. The wo s a egies a e e alua ed quan i a i ely wi h one eal LIDAR da ase and wo RGBD da ase s (one eal and he o he syn he ic). This e al- ua ion allows o objec i ely compa e he wo sys ems. The bes sys em is used o ob ain 3D econs uc ions e en wi hou pho ome ic images, jus wi h a LIDAR senso de eloped a Beamagine (a spin-o o UPC de eloping LIDARs based on p op ie a y echnology). The pape is o ganized as ollows. Sec ion 2 e iews he s a e o he a in 3D SLAM sys ems. Sec ion 3 explains he adap a ions o SLAM sys ems o LIDAR da a. Sec ions 4 and 5 p o ide e alua ion esul s and conclusions. 2. STATE OF THE ART The basics o SLAM sys ems capable o c ea ing h ee dimen- sional maps we e in es iga ed in he o m o 3D g ids [2] and 3d geome ic ea u es [3]. The i s 3D SLAM sys ems used mono came as [4], s e eo came as [5] o 3D LIDARs [6]. Mo e ecen ly, he a ailabili y o eal- ime dense dep h sen- so s (RGBD) has eased he li e econs uc ion o eal scenes. Mic oso de elops Kinec Fusion [7] in 2011, an algo i hm allowing 3D econs uc ions a 30 ps aking ad an age o he ecen ly launched Kinec ma icial dep h senso . One yea la e , PCL [8] inco po a es a simila open-sou ce ool known as KinFu [9]. Bo h sys ems use a oxelized ep esen a ion o he scene named TSDF model (T unca ed Signed Dis ance Func ion model [10]), whe e each oxel s o es he dis ance o he closes su ace and a con idence weigh . The main limi a- ion o hese sys ems is he inabili y o map a eas la ge han he model. This limi a ion was elimina ed a he same ime by Kin inuous [11] and KinFu la ge-scale [12]. Kin inuous implemen s an unbounded mapping o he en i- onmen on op o KinFu. P ecisely, i inco po a es he abil- i y o i ually ansla e he TSDF model when new es ima ed came a poses exceed a dimension independen h eshold. Kin- inuous was imp o ed o be mo e obus agains challenging scenes [13], such as la ge came a displacemen s o lack o 3D dep h ea u es, while also aiming o elimina e he accumu- la ed d i o p e iously egis e ed ames [14]. The d i elim- ina ion is known as loop closu e. I happens when he sen- so e isi s a p e ious loca ion by op imizing all he a ec ed ans o ma ions wi h a pose op imize (iSAM) and a non- igid me hod ha co ec s he econs uc ion. KinFu la ge-scale is now a simple ool simila o he o iginal Kin inuous wi hou he eal- ime map ex ac ion. RGB-D SLAM [15] is ano he eal- ime sys em wi h Robo Ope a ing Sys em (ROS) suppo [16]. The ans o ma ions be ween poses a e ob ained by de ec ing key-poin s o incom- ing ames, compu ing ea u es and inding co espondences wi h olde ones. The sys em also does loop closu e wi h a pose op imize (g2o). The Oc oMap amewo k is used o c e- a e econs uc ions using he op imized ajec o y. The sys em was imp o ed [17] and now includes: a beam-based en i on- men measu emen model (EMM) ha alida es he es ima ed ans o ma ions acco ding o occlusion p obabili ies, a selec- ion s a egy o candida e ames o compa isons based on explo ing he geodesic g aph neighbo hood o he p e ious ame, and he use o key- ames o simpli y he sea ch. Elas icFusion [18] is ano he eal- ime sys em de eloped by some o he au ho s o Kin inuous. I is based on a su ace model ins ead o TSDF, he loop closu e is done wi hou a pose op imize by non- igidly de o ming he a ec ed su aces. RTAB-Map is ano he eal- ime sys em wi h ROS suppo ha can wo k wi h 2D LIDARs and s e eo se ups (apa om RGBD came as). I is based on a g aph o links and nodes. The nodes con ain in o ma ion abou he poses o he obo and he links s o e igid ans o ma ions be ween nodes. The ans o ma ions a e ob ained using 3D isual wo ds co e- spondences and main ained wi h TORO (T ee-based ne wORk Op imize ) allowing o p opaga e he e o h ough links a e loop closu es. Addi ionally, RTAB-Map inco po a es a p ox- imi y module o ind loop closu es wi h 2D LIDARs ha helps in si ua ions when he RGBD came as do no ha e enough in o ma ion. The s ong poin o RTAB-Map is a memo y- e icien loop closu e de ec ion app oach. 3. LIDAR ADAPTATION F om he SLAM s a egies explo ed in he p e ious sec ion, we ha e selec ed bo h Kin inuous and RTAB-Map (a ailable on Gi hub) o wo k wi h LIDAR da a. The easons a e ha Kin inuous is supposed o pe o m be e han Elas icFusion wi h noisy LIDAR da a and ha RTAB-Map is expec ed o imp o e RGB-D SLAM wi h LIDAR, since he EMM o RGB- D SLAM assumes dense dep h measu emen s, and he loop closu e app oach o RTAB-Map seems o be mo e e icien . We ha e adap ed he SLAM algo i hms o LIDAR da a, and we desc ibe he adap a ions acco ding o he speci ic senso se up o he LIDAR da ase and Beamagine da a. 3.1. Adap a ion o LIDAR da ase The KITTI da ase [19] is he one chosen o he adap a ion o SLAM algo i hms o LIDAR da a as i allows quan i a i e e alua ion. I consis s o 22 sequences abou di e se a ic en i onmen s (highway, u al and ci y). Rega ding he sen- Fig. 1: Dep h image ob ained a e he p ojec ion o a KITTI LIDAR scan, wi h i s co esponding colo image below. Pix- els wi hou dep h alues in he LIDAR scan a e colo ed in blue in he dep h image o ease isualiza ion. so se up, i is composed o : 2x g ay-scale and colo cam- e as, 1x o a ing 3D LIDAR and 1x ine ial and GPS uni . In his adap a ion, we only use he images o he le colo cam- e a and he scans o he LIDAR om he al eady ec i ied, undis o ed and synch onized e sion o he da ase . Bo h he p ojec ion ans o ma ion o he ec i ied came as and he ex- insic ans o ma ion om 3D LIDAR coo dina es o came a coo dina es a e p o ided in [19]. In he i s pa o he adap a ion, we con e ed he ele en KITTI sequences wi h a ailable g ound- u h o he PNG o - ma o he RGB-D SLAM da ase [20]. This was done wi h a ool ha basically p ojec s he 3D LIDAR scans o he se- lec ed came a (in ou case he le one wi h colo ) and cal- cula es i s dep h alues. Then, e e y alue is quan ized o a 16-bi unsigned ep esen a ion conside ing he maximum LI- DAR ange (120 me e s). The esul ing quan iza ion s ep is much lowe han he one p o ided by he manu ac u e (1.8  20 millime e s). Gi en he LIDAR p ope ies and he came a FOVs, only abou 32% o he poin s o a comple e scene a e p ojec ed o he came a plane wi hin he sequence dependen image size, whe e hal o hese poin s a e on -p ojec ed. A esul o he dep h p ojec ion om a LIDAR scan is shown along wi h he co esponding colo ame in igu e 1. The emaining pa s o he adap a ion speci ic o each sys em a e desc ibed below. 3.1.1. Kin inuous applied o KITTI da a The implemen a ion o Kin inuous uses log iles in KLG o - ma as inpu . This o ma consis s o s o ing in a single ile all he in o ma ion o a sequence: he imes amps and a com- p essed e sion o he dep h and colo images. The main au- ho o Kin inuous p o ides some ools o c ea e KLG log iles di ec ly om da a-s eams o senso s like Kinec and X ion P o Li e. Tha said, a con e sion om PNG RGB- D SLAM o ma o KLG o ma was needed. Fo una ely, he implemen a ion o his con e sion was al eady done in a Gi Hub eposi o y [21]. This eposi o y con ains a ool called png o klg ha essen ially c ea es a KLG log ile om he ame pai s p o ided by an associa ions ex ile, con e s he imes amps om seconds o mic o-seconds and he scaled dep h measu emen s o millime e uni s. The co e o Kin inous is he cubic TSDF model ha , in i s de aul con igu a ion, has a side leng h in oxels o 512 and a eal wo ld equi alence o 6 me e s. The con e ed LIDAR da a has a heo e ical maximum ange o 120 me e s. These wo s a emen s make he sys em and he da a incompa ible. The only wo ways o sol e his is by adap ing he da a o Kin- inuous o Kin inuous o he da a. Rega ding he sys em adap- a ion, inc easing he numbe o oxels o he cubic TSDF model may be an op ion, bu i equi es a complex code mod- i ica ion and is expec ed o ail due o low densi y o poin s wi hin he model. This lack o poin s would be p oduced by he low numbe o LIDAR poin s (100K pe scan) and he low scan- a e (10 Hz) ela ed o he a e age LIDAR mo emen (ca mo ion). On he o he hand, he da a adap a ion could be achie ed ei he by inc easing he eal wo ld equi alence o a single oxel o scaling he eal wo ld dimensions, bo h o hem a he cos o losing p ecision. The second op ion was chosen, and implemen ed by scaling he da ase wi h a wo ld scale ac o , which was implici ly in oduced along wi h he dep h quan iza ion ac o in he png o klg ool. This allows he gene a ion o KLG iles wi h di e en wo ld scale ac o s. A e some es ing, he de ini i e wo ld scale ac o was se o 20 (1 scaled me e o he algo i hm co esponds o 20 wo ld me e s). This comes om he ac ha he ac ual maximum dep h o LIDAR scans was abou 80 m. and he dep h limi ha Kin inuous implemen a ion allows o p ojec is 4m. (as i conside s ha la ge Kinec dep hs a e oo noisy). When execu ing Kin inuous, we se he shi ing h eshold o 16 oxels (maximum acco ding o he au ho ) since he dis- ance a eled by he came a a di e en ames is la ge , due o high eloci y (ca in KITTI s hand-held came as in RGB- D SLAM da a) and low ame a e (30 s 10 ps). Also, he pa ame e subsample pose g aph was deac i a ed o expo all op imized poses o he g aph when loop closu e is enabled. 3.1.2. RTAB-Map applied o KITTI da a The RTAB-Map implemen a ion uses images s o ed in egula iles as inpu , hus no equi ing he png o klg ool. The im- ages need o be al eady associa ed in disc since i does no ac- cep an associa ions ex ile as synch oniza ion in o ma ion. Luckily, he adminis a o o RTAB-Map al eady p o ides a modi ied e sion o he RGB-D SLAM associa ions ool ha , ins ead o expo ing he pai ing in o ma ion in a ex ile, c e- a es di ec o ies and mo es he synch onized images esul ing om he associa ion p ocess. In his case, he wo ld scale was unnecessa y since RTAB- Map sys em is no es ic ed o he anges o s uc u ed ligh senso s. Ne e heless, we decided o execu e bo h e sions, hus allowing o e i y he co ec implemen a ion o he wo ld scaling ac o and i s e ec . We had o locally modi y he RTAB-Map implemen a ion o allow o a dep h scale ac- o lowe han 1 s ep/millime e which was he case when using he con e ed e sion o he da ase wi h ki y o png. We used he RGB-D da ase command-line ool o execu - ing wi h RTAB-Map ins ead o he GUI in e ace. This ool sa es a SQLi e da abase wi h all he in o ma ion ela ed o he SLAM and expo s he poses in he selec ed o ma . The maps can be c ea ed wi hou RTAB-Map GUI using he expo ex- ample (a ailable in he examples olde o he Gi hub epos- i o y) implemen ed by one au ho ( hanks o Ma hieu Labb´ e) a e asking a ques ion in he o icial RTAB-Map o um. 3.2. Adap a ion o Beamagine da a Unlike he KITTI da ase , Beamagine da a comes om a sin- gle senso : a 3D LIDAR wi h i s in a ed ligh based ange measu emen s, wi hou a egis e ed colo came a. This LI- DAR, di e en om he one in KITTI, is s a ic and on - acing, and i s main speci ica ions a e 5 Hz, 0.5Mpoin s/s, FOV: 54.5◦h, 20◦ , ange 165m., 200x600 sampling poin s. The lack o a pho ome ic came a in he Beamagine se -up opens he challenge o es SLAM sys ems wi hou exploi ing pho ome ic RGB da a in a dense, egula ly sampled as e image. Visual SLAM de ec s singula image poin s o ind co espondences be ween ames o be egis e ed. A his poin , we p opose o eplace he dense pho ome ic in o ma- ion by he in a ed in ensi y o he LIDAR poin s. This idea was implemen ed in a ool called beamagine o klg simila o he one implemen ed o KITTI. Basically, i eads he LI- DAR scans (s o ed in sepa a ed pcd iles), con e s he me ic uni s om millime e s o me e s, p ojec s he poin s o a sim- ula ed came a plane and calcula es hei associa ed dep h and in ensi y alues. As came a pa ame e s, we only used a ocal alue o each dimension ( o accoun o pe spec i e p ojec- ion) and he image size, since no in insic LIDAR calib a- ion was a ailable. The image size was selec ed simula ing he highes sampling equency in each dimension. I was di ec ly se o 200 pixels e ical, and 1364 pixels ho izon- al, since he o -cen e poin s ha e highe esolu ion han he cen e ones (in angula measu es: 0.04◦ s. 0.15◦) due o de- sign cons ain s o he senso . Then, he wo ocal pa ame e s we e ob ained conside ing bo h he image dimensions and he wo LIDAR FOVs. Gi en he simula ed came a pa ame e s, abou 99% o poin s a e co ec ly p ojec ed. A e p ojec ion, each alue is quan ized o unsigned 16-bi s conside ing he dynamic ange o he measu e. The dep h quan iza ion s ep is 2,5 mm (a bi la ge han KITTI’s), bu again i is conside ed o be su icien . A esul o he dep h calcula ion is shown in igu e 2. Fig. 2: Dep h image ob ained om he p ojec ion o he poin cloud o a Beamagine LIDAR scan. Pixels wi hou dep h al- ues a e se o blue in he dep h image o ease isualiza ion. The pho o below was aken some days a e he cap u es om a simila poin o iew. Fig. 3: In a ed alues om a Beamagine scan p ojec ed in an image plane be o e and a e in e pola ion. Pixels wi h no alue a e se o blue in he uppe image o ease isualiza ion. The con en wi hin he ed boxes co esponds o he same im- ages wi hou elimina ing he sa u a ed and null alues o he dis ibu ion ha p oduced high spa ial equencies. Fu he mo e, in o de o ob ain in a ed da a simila o a dense as e image (wi hou holes), he gene a ed in ensi y images we e pos -p ocessed wi h an inpain ing s age. This was done o ul ill he SLAM sys ems equi emen s and o ease he de ec ion o singula poin s o co espondences. Fo he inpain ing, we used he 8- connec i i y e sion o a mo - phological in e pola ion echnique [22], ha p ese es he o ig- inal in a ed alues o he p ojec ed poin s and i s ansi ions, hanks o he use o geodesic dis ance. This echnique is e - icien ly implemen ed as an i e a i e p ocess: i s , he se o ini ial pixels a e p opaga ed using geodesic dila ion and, hen, he ansi ions gene a ed a e eco e ed by applying he mo phological Laplacian, whe e he a e age alues a e used along wi h he o iginally p ojec ed ones o p opaga ion in subsequen i e a ions. A e applying he inpain ing s age, some senso noise wi h a high spa ial equency be ween he sa u a ed and null alues was disco e ed in he in a ed al- ues o he Beamagine senso . The noise was elimina ed by disca ding he wo his og am peaks which implied o end up using abou 30% o he dynamic ange. In o de o e alua e he e ec o using he in ensi y o he poin s ins ead o he colo images, he same adap a ion was done in he png o klg ool o he KITTI da ase . Figu es 3 and 4 show a ame and i s in e pola ed e sion o Beamag- ine and KITTI da a, espec i ely. Fig. 4: In a ed image be o e and a e in e pola ion om a KITTI LIDAR scan. The speci ic adap a ion o each SLAM sys em is simila o he one explained in sec ion 3.1 since he gene a ed da a o ma is he same. 4. RESULTS In his sec ion, we p esen and discuss quan i a i e esul s o e alua ing he sys ems wi h he LIDAR adap a ion men ioned in sec ion 3.1, and he e alua ion done wi h RGBD da ase s bo h na u al and syn he ic. Quali a i e esul s ob ained o he LIDAR adap a ion o sec ion 3.2 a e also shown. 4.1. T ajec o y e alua ion The quan i a i e ajec o y e alua ion was done wi h he Ab- solu e T ajec o y E o me ic (ATE) [23] o bo h eal Kinec and LIDAR da a modali ies using g ound- u h a ailable in RGB-D SLAM and KITTI da ase s. 4.1.1. Es ima ed T ajec o ies on he RGB-D SLAM da ase Depending on he sys em, we pe o med di e en es s swi ch- ing he loop closu e componen and a ying pa ame e s o SLAM algo i hms. Fo Kin inuous, we ied all possible cos - combina ions excep FOVIS alone ( ha could no be se ). The combina ions a e: ICP,RGB-D,FOVIS/ICP,FOVIS/RGB-D, FOVIS/ICP+RGB-D and ICP+RGB-D. Fo RTAB-Map, we ied o modi y i s key-poin and ea u e desc ip o ex ac- o s, always using ame- o-map odome y and 3D o 3D mo- ion es ima ion. This includes: su ,si ,o b,kaze,b isk, g /b ie ,g /o b,g / eak, as / eak, as /b ie ,su /b ie and su /o b. The de aul beha io o RTAB-Map when i canno com- pu e a ans o ma ion (minimum o 20 inlie s by de aul ) o an incoming ame is o disca d i . Con e sely, he de aul beha io o Kin inuous is o epea he las pose. This ac , ende s he ajec o y e alua ion o bo h sys ems somewha Sequence Kin inuous RTAB-Map desk 0,052 0,082 oom 0,224 0,128 desk2 0,073 0,045 la ge no loop 0,465 0,332 pionee slam2 2,186 - long o ice household 0,048 0,037 AVERAGE 0,172 0,125 Table 1: Bes RMSE esul s o ATE [23] on he Kinec RGB- D SLAM da ase (bes esul s pe sequence in bold). biased. The way we app oached a ai compa ison was by ex- ecu ing RTAB-Map in a ixed and an adap i e o m, and using only he ixed o m o he compa ison. The ixed o m con- sis s o se ing a maximum inlie dis ance o he ea u e co e- spondences o a ixed alue o all he sequences and disca d- ing he execu ions whe e he sys em is no able o compu e he ans o ma ion o any o he sequence ames. The same was applied o he execu ions whe e Kin inuous ou pu s epea ed ans o ma ions. The adap i e o m consis s o s a ing wi h a low inlie dis ance and, i a any ame o he sequence he ans o ma ion canno be compu ed, he inlie dis ance alue is inc eased by a ac o and s a s again, un il success o un- il eaching a maximum alue. While his la e o m ends o gi e mo e accu a e esul s, i is sequence dependen and would no be applicable o eal- ime si ua ions. F om all he e alua ions we un, we picked he bes pe - o ming combina ion o pa ame e s o each sys em, based on he a e age RMSE o he sequences. Fo Kin inuous, he bes combina ion was ICP+RGB-D, while o RTAB-Map, i was a unned e sion o he g /b ie combina ion. Loop closu e was enabled in bo h cases. Speci ically, he pa ame e s modi- ied whe e he quali y le el o he g (good ea u es o ack) key-poin ex ac o , ha was se o 0.005, and he minimum Euclidean dis ance be ween de ec ed co ne s, se o 5 pixels. The esul s a e shown in able 1, whe e RTAB-Map pe o ms abou 40% be e in a e age han Kin inuous in ajec o y es- ima ion. And his happens consis en ly in all sequences bu o he pionee slam 2 sequence, whe e i is no able o com- pu e all he ans o ma ions when he inlie dis ance is ixed a 0.1 me e s. This ac happens wi h all es ed combina ions and only yields esul s when RTAB Map is execu ed wi h he adap i e modali y ha allows o a g ea e inlie dis ance. Also no e ha , in his sequence, Kin inuous is able o com- pu e he ajec o y bu wi h an a e age RMSE o abou 2m, wi h alues in a ange o [0,196m, 3,614m]. No e ha RTAB- Map is mo e accu a e han Kin inuous o abou one ou h o he ajec o y leng h, whils , in he emainde , Kin inuous main ains i s pe o mance while RTAB-Map d i s. 4.1.2. Es ima ed T ajec o ies on he KITTI da ase Unlike o he p e ious da ase , he e we e alua ed he ajec- o y o all en sequences wi h a ailable g ound- u h. Fo RTAB-Map, we swi ched again he loop closu e componen bu only conside ing he combina ion ha ga e bes esul s (g /b ie ) in he Kinec ajec o y baseline. The quali y le el o he g key-poin ex ac o was changed om 0,005 o 0,0005 and he minimum Euclidean dis ance be ween de ec ed co ne s was inc eased by one pixel. Again, we execu ed all cos -combina ions o Kin inuous. In he e alua ion o his LIDAR da ase , he bes cos - combina ion o Kin inuous was he RGB-D independen one, whe he execu ed wi h o wi hou loop closu e, while he one o RTAB-Map is wi h loop closu e enabled. Table 2 com- pa es hese esul s. No e ha RTAB-Map is abou 5 imes be - e han Kin inuous in a e age in ajec o y es ima ion. Apa om his, in igu e 5 we show a plo om sequence 07 com- pa ing he ansla ional pa o he es ima ed ajec o y wi h i s co esponding g ound- u h. The plo isually p o es ha he ajec o y is be e es ima ed by RTAB-Map han by Kin i- nuous. The p ojec ion does no allow o isualize he e ical componen o he di e ences. 4.2. E alua ion o he 3D econs uc ed map Fo he quan i a i e e alua ion o he 3D mapping gene a ion unc ionali y o SLAM algo i hms we ha e chosen o use he ool p o ided by he main au ho o Kin inuous. This ool compu es as me ic he poin - o-poin dis ance be ween he g ound- u h and he es ima ed maps on a syn he ic da ase o li ing- oom sequences known as ICL-NUIM [24]. 4.2.1. E alua ion o he mapping o he ICL-NUIM da ase All he ou li ing- oom sequences l k 0..3 we e used wi h and wi hou simula ed Kinec noise. Fo RTAB-Map, we used he same bes combina ion ound in he ajec o y baseline o sec ion 4.1.1, wi h and wi hou he loop closu e componen . Rega ding he RTAB-Map econs uc ion ex ac ion, and in o de o pe o m a compa ison, a oxel g id il e wi h he same lea size as he one used by Kin inuous (6/512) was used. Fo he c ea ion o he poin clouds, he same maximum leng h as Kin inuous is used (4 me e s) and a decima ion in he colo image by a ac o o 9 was applied o ob ain a simila numbe o poin s o he maps o bo h sys ems. Addi ionally, o he e sion o he da ase wi h noise, a local smoo hing il e was ied in RTAB-Map bu , as he compu a ional ime o ex ac ion inc eased and some ine walls o he econs uc- ions we e il e ed be o e he emo al o some noisy pa s, i was no included o he compa ison. Again, o Kin inuous, all cos -combina ions we e ied, il e ing he noisy ex ac ed poin s om he ze o c ossing su ace o he slices wi h a min- imum oxel weigh h eshold o 8 (de aul ). Sequence Kin inuous Kin inuous, LC RTAB-Map RTAB-Map, LC 00 149,7 149,7 30,9 11,5 01 488,6 489,1 - - 02 289,8 289,8 34,5 29,0 03 2,3 2,3 6,9 7,3 04 11,8 11,9 11,7 11,7 05 93,3 93,4 21,7 18,5 06 203,8 203,7 - - 07 21,9 21,9 3,2 2,2 08 65,1 65,1 30,0 26,6 09 77,1 77,2 17,9 15,7 10 38,0 38,1 8,8 8,5 AVERAGE 83,2 83,3 18,4 14,6 Table 2: Bes RMSE esul s o ATE [23], wi h and wi hou loop closu e (LC), on he KITTI LIDAR da ase (bes esul s pe sequence in bold). (a) (b) Fig. 5: Di e ences be ween he es ima ed ajec o ies o sequence 07 (in g een) and he g ound u h (in black) p ojec ed o he xz plane: (a) Kin inuous, (b) RTAB-Map. Table 3 summa izes he esul s o bo h da a modali ies (wi h and wi hou noise). The bes esul s o Kin inuous a e ob ained wi h ICP cos and wi h loop closu e. Howe e , he sequence l k 0 is conside ed wi hou loop closu e, since he de o ma ion g aph ailed wi hou sa ing any esul o Kin i- nuous in he o iginal e sion o he sequence, and o RTAB- Map i imp o ed mo e han double wi hou loop closu e. Sim- ila ly, o RTAB-Map all esul s a e picked wi h loop closu e excep o he i s sequence. 4.3. Recons uc ions As inal quali a i e esul s o his sec ion, we p esen he ob- ained RTAB-Map econs uc ions wi h 3D LIDAR da a o bo h he KITTI and Beamagine scaled da ase s. Sequence Kin inuous RTAB-Map Kin inuous RTAB-Map Modali y O iginal O iginal Noise Noise AVG poin s 471K 555K 441K 863K l k 0 4,4 12,7 6,4 50,2 l k 1 5,6 4,7 8,9 69,9 l k 2 4,3 7,8 9,0 51,0 l k 3 74,2 6,3 77,2 58,5 Table 3: RMS poin - o-poin dis ance o e alua ion o he econs uc ed 3D map in he ICL-NUIM da ase (in bold, bes echnique esul s o each modali y). 4.3.1. 3D econs uc ion o he KITTI da ase Fo his da ase 3D econs uc ions a e gene a ed o he ele en sequences e alua ed in sec ion 4.1.2. The expo ool men- ioned in sec ion 3.1.2 is used in place o he RTAB-Map wi h (a) (b) Fig. 6: RTAB-Map econs uc ions o sequence 07 unscaled, om a simila poin o iew o he one o he in ensi y, dep h and colo ames showed in sec ion 3.1, igu e 1. The econ- s uc ions modes a e: (a) Mesh (b) Poin Cloud. Fig. 7: RTAB-Map econs uc ion o sequence 08 unscaled. The snapsho on he le shows a bike in on o he ca , wi h he 3D o e all econs uc ion o he ajec o y on he igh . The cen al image shows a zoom in on he ed ec angle, wi h he da ke ace o he mo ing bike clea ly isible. GUI ins alla ion, ha ing as inpu he da abases gene a ed wi h he same con igu a ion ha p oduced he compa ed ajec o y esul s. As a eminde , o hose compa isons, we used he le colo came a and he 3D LIDAR o he ca senso se up. Due o he la ge numbe o econs uc ions and he di icul y o showing he 3D econs uc ions in a pape epo , we only show a ew o hem. Fo example, igu e 6 shows a de ail o he econs uc ion o sequence 07. Sequence 08 is one o he mos complex and la ge. A op iew o i s 3D econs uc ion is shown in igu e 7. As men ioned in sec ion 3.2, we also ied o disca d he pho ome ic in o ma ion and use only he LIDAR da a p o- ided in he KITTI da ase . Un o una ely, we could no ob- ain any good econs uc ion a he ime o w i ing his epo . 4.3.2. 3D econs uc ion o he Beamagine da a In his case, we used he adap a ion desc ibed in sec ion 3.2 wi h a da ase o 8 sequences, whe e each one con ains a hun- d ed ames. As a eminde , in hese sequences, we only had da a coming om he 3D LIDAR. In spi e o his si ua ion, we we e able o ob ain some econs uc ion esul s. Fo ins ance, igu e 8a shows pa o a econs uc ion ha co esponds o he pho o on he side (8b). The pho os we e aken some days a e he da ase cap u e om a simila poin o iew. Also, he poin o iew is simila o he one o he dep h and in en- si y ames om igu es 2 and 3. (a) (b) Fig. 8: RTAB-Map unscaled econs uc ions, (a) Mesh econ- s uc ed om LIDAR in a ed and dep h da a, and (b) Pho o aken some days a e he cap u e ( o compa ison pu poses). 5. CONCLUSIONS We ha e success ully adap ed wo SLAM sys ems (Kin inuous and RTAB-Map) o wo k wi h LIDAR da a. We ha e ob ained a quan i a i e baseline wi h indoo RGBD da a by e alua ing he mapping ( econs uc ion) and loca ion ( ajec o y) pe o - mance o bo h sys ems. Besides his, we ha e ca ied ou a ajec o y e alua ion wi h ou doo LIDAR da a. All hese ob- jec i e e alua ions ha e been pe o med on publicly a ailable da ase s wi h anno a ed g ound- u h. Addi ionally, we ha e es ed he bes sys em in a LIDAR da ase lacking a isible colo came a, hus only exploi ing he me ic in o ma ion o he LIDAR and he in a ed al- ues o he p ojec ed scan poin s. We p opose an in e pola- ion me hod o he emp y a eas o allow o ea u e de ec o s needed by SLAM algo i hms o co espondence ma ching. Wi h his challenging da a we ha e ob ained some econs uc- ions om he s ee s o Te assa, a ci y nea Ba celona whe e UPC has one o i s campuses. We would like o highligh he ollowing poin s esul ing om ou explo a ion: •In indoo eal scena ios RTAB-Map is sligh ly be e han Kin inuous o ajec o y es ima ion. Howe e , based on poin - o-poin map di e ences, Kin inuous is be e in 3D econs uc ion o syn he ic indoo s wi h simula ed Kinec noise, p obably hanks o i s TSDF model. Fo ou doo eal da a RTAB-Map undoub edly pe o ms be e based on ATE. •Wi h he scan p ojec ions done in he KITTI da ase abou 84% o he a ailable 3D poin s a e los , hence, he ob ained econs uc ions ha e low densi y o poin s in he pa s ha a e no cap u ed by he came a FOV. •Fo da a cap u ed wi h less han 6DoF (like KITTI), he cubic shape o he TSDF olume (used in Kin inuous) is a was e o esou ces, since a la ge pa o i is ne e used. •The use o a spa sely sampled in a ed image in place o a high esolu ion isible image makes he SLAM p oblem mo e di icul , bu simpli ies he senso se up. •Dynamic mo emen s o objec s b eak he assump ion o a s a ic wo ld p oducing duplica ions in he map. As u u e wo k, we would like o con inue wi h: a p ecise in insic calib a ion o he Beamagine LIDAR, a egis a ion o a hi- es colo came a wi h he Beamagine LIDAR da a, and exploi ing he ad an ages o he eal- ime ROS w appe o RTAB-Map. 6. REFERENCES [1] P. F. McManamon, “Re iew o lada : a his o ic, ye eme ging, senso echnology wi h ich phenomenol- ogy,” Op ical Enginee ing, ol. 51, no. 6, 2012. [2] H. Mo a ec, “Robo spa ial pe cep ion by s e eoscopic ision and 3d e idence g ids,” Pe cep ion, 1996. [3] N. Ayache and O. D. Fauge as, “Building, egis a ing, and using noisy isual maps,” The In e na ional Jou - nal o Robo ics Resea ch, ol. 7, no. 6, pp. 45–65, 1988. [4] E. Eade and T. D ummond, “Scalable monocula slam,” in Compu e Vision and Pa e n Recogni ion, ol. 1. IEEE Compu e Socie y, 2006, pp. 469–476. [5] M. A. Ga cia and A. Solanas, “3d simul aneous localiza- ion and modeling om s e eo ision,” in Robo ics and Au oma ion ICRA’04, ol. 1. IEEE, 2004, pp. 847–853. [6] H. Su mann, A. N¨ uch e , and J. He zbe g, “An au- onomous mobile obo wi h a 3d lase ange inde o 3d explo a ion and digi aliza ion o indoo en i on- men s,” Robo ics and Au onomous Sys ems, ol. 45, no. 3-4, pp. 181–198, 2003. [7] R. A. Newcombe, S. Izadi, O. Hilliges, D. Molyneaux, D. Kim, A. J. Da ison, P. Kohi, J. Sho on, S. Hodges, and A. Fi zgibbon, “Kinec usion: Real- ime dense su - ace mapping and acking,” in IEEE In l. Symposium on Mixed and Augmen ed Reali y, Oc 2011, pp. 127–136. [8] “Poin cloud lib a y.” [Online]. A ailable: h p: //poin clouds.o g/ [9] M. Pi o ano, “Kin u - an open sou ce implemen a ion o Kinec Fusion+ case s udy: implemen ing a 3d scanne wi h PCL,” UniMi, Tech. Rep., 2012. [10] B. Cu less and M. Le oy, “A olume ic me hod o building complex models om ange images,” in 23 d Annual Con e ence on Compu e G aphics and In e ac- i e Techniques. New Yo k: ACM, 1996, pp. 303–312. [11] T. Whelan, M. Kaess, and M. Fallon, “Kin inuous: Spa- ially ex ended kinec usion,” RSS Wo kshop on RGB-D: Ad anced Reasoning wi h Dep h Came as, p. 7, 2012. [12] F. He edia and R. Fa ie , “Kinec Fusion ex ensions o la ge scale en i onmen s,” 2012. [Online]. A ailable: h p://www.poin clouds.o g/blog/s cs/ he edia/ [13] T. Whelan, H. Johannsson, M. Kaess, J. J. Leona d, and J. McDonald, “Robus eal- ime isual odome y o dense RGB-D mapping,” IEEE In l. Con e ence on Robo ics and Au oma ion, pp. 5724–5731, 2013. [14] T. Whelan, M. Kaess, J. J. Leona d, and J. McDonald, “De o ma ion-based loop closu e o la ge scale dense RGB-D SLAM,” IEEE/RSJ In l. Con e ence on In elli- gen Robo s and Sys ems, pp. 548–555, 2013. [15] F. End es, J. Hess, N. Engelha d, and J. S u m, “An e alua ion o he RGB-D SLAM sys em,” ICRA, ol. 3, no. c, pp. 1691–1696, 2012. [16] “ROS.o g |Powe ing he wo ld’s obo s.” [Online]. A ailable: h p://www. os.o g/ [17] F. End es, J. Hess, J. S u m, D. C eme s, and W. Bu - ga d, “3-D Mapping Wi h an RGB-D Came a,” IEEE T ans. on Robo ics, ol. 30, no. 1, pp. 177–187, 2014. [18] T. Whelan, S. Leu enegge , R. Salas Mo eno, B. Glocke , and A. Da ison, “Elas icFusion: Dense SLAM Wi hou A Pose G aph,” Robo ics: Science and Sys ems XI, 2015. [19] A. Geige , P. Lenz, C. S ille , and R. U asun, “Vision mee s obo ics: The KITTI da ase ,” Jou nal o Robo ics Resea ch, ol. 32, no. 11, pp. 1231–1237, 2013. [20] J. S u m, N. Engelha d, F. End es, W. Bu ga d, and D. C eme s, “A benchma k o he e alua ion o RGB- D SLAM sys ems,” IEEE In e na ional Con e ence on In elligen Robo s and Sys ems, pp. 573–580, 2012. [21] L. Jacky, “png o klg.” [Online]. A ailable: h ps: //gi hub.com/HTLi e/png o klg [22] J. R. Casas, P. Salembie , and L. To es, “Mo phological in e pola ion o ex u e coding,” IEEE In l. Con e ence on Image P ocessing, ol. 1, pp. 526–529, 1996. [23] J. S u m, N. Engelha d, F. End es, W. Bu ga d, and D. C eme s, “A benchma k o he e alua ion o RGB- D SLAM sys ems,” IEEE In e na ional Con e ence on In elligen Robo s and Sys ems, pp. 573–580, 2012. [24] A. Handa, T. Whelan, J. McDonald, and A. J. Da ison, “A benchma k o RGB-D isual odome y, 3D econ- s uc ion and SLAM,” IEEE In e na ional Con e ence on Robo ics and Au oma ion, pp. 1524–1531, 2014.