Brno Urban Dataset: Winter Extension

Da a in B ie 40 (2022) 107667 Con en s lis s a ailable a ScienceDi ec Da a in B ie jou nal homepage: www.else ie .com/loca e/dib Da a A icle B no u ban da ase : Win e ex ension Adam Ligocki ∗, Ales Jelinek, Ludek Zalud B no Uni e si y o Technology, Technicka 12, B no 612 00, Czechia a i c l e i n o A icle his o y: Recei ed 8 Sep embe 2021 Re ised 21 Oc obe 2021 Accep ed 30 No embe 2021 A ailable online 3 Decembe 2021 Keywo ds: Mul imodal da ase Na iga ion da a RGB came a IR came a 3D LiDAR RTK GNSS IMU Neu al ne wo ks a b s a c This pape p esen s ou la es ex ension o he B no U ban Da ase (BUD), he Win e Ex ension (WE). The da ase con- ains da a om commonly used senso s in he au omo i e in- dus y, like ou RGB and single IR came as, h ee 3D LiDARs, di e en ial RTK GNSS ecei e wi h heading es ima ion, he IMU and FMCW ada . Da a om all senso s a e p ecisely imes amped o u u e offline in e p e a ion and da a u- sion. The mos significan gain o he da ase is he ocus on he win e condi ions in snow-co e ed en i onmen s. Only a ew public da ase s deal wi h hese kinds o condi ions. We eco ded he da ase du ing Feb ua y 2021 in B no, Czechia, when esh snow co e s he en i e ci y and he su ound- ing coun yside. The da ase con ains si ua ions om he ci y cen e , subu bs, highways as well as he coun yside. O e all, he new ex ension adds h ee hou s o eal-li e affic si u- a ions om he mid-size ci y o he exis ing 10 h o o igi- nal eco ds. Addi ionally, we p o ide he p ecalcula ed YOLO neu al ne wo k objec de ec ion anno a ions o all fi e cam- e as o he en i e old da a and he new ones. The da ase is sui able o de eloping mapping and na iga ion algo i hms as well as he collision and objec de ec ion pipelines. The en i e da ase is a ailable as open-sou ce unde he MIT license. ©2021 The Au ho s. Published by Else ie Inc. This is an open access a icle unde he CC BY license ( h p://c ea i ecommons.o g/licenses/by/4.0/ ) ∗Co esponding au ho . E-mail add ess: adam.ligocki@ u b .cz (A. Ligocki). h ps://doi.o g/10.1016/j.dib.2021.107667 2352-3409/© 2021 The Au ho s. Published by Else ie Inc. This is an open access a icle unde he CC BY license ( h p://c ea i ecommons.o g/licenses/by/4.0/ ) 2 A. Ligocki, A. Jelinek and L. Zalud / Da a in B ie 40 (2022) 107667 Specifica ions Table Subjec Compu e Science, A ificial In elligence Specific subjec a ea I is a da ase o mapping and au onomous agen de elopmen . Type o da a RGB and IR images, poin cloud scans, ex da a logs How da a we e acqui ed Real-li e affic eco ding, using he senso y amewo k moun ed on he oo o he ca . The amewo k employs ou 1920 ×1200px@10 ps RGB came as, single he mal (IR) came a 640 ×512px@30 ps, wo Velodyne HDL-32E LiDARs, a single Li ox Ho izon one, BX982 RTK GNSS ecei e wi h di e en ial an ennas, Xsens MTI-G-700 IMU, and he TI mmWa e 1648 FMCW ada . Da a o ma Raw senso da a a e s o ed in h ee o ms. The came a da a a e s o ed as an mp4, h265 ideo. The LiDAR s ans a e s o ed as .pcd scan files. The e is a co esponding ex file o bo h da a ypes ha desc ibes when he da a we e cap u ed. The es o he da a a e s o ed as a CSV eco d ha desc ibes he imes amp and he measu ed alues. Addi ionally, we spli he da ase in o se e al eco ding sessions. Each session akes place in a dedica ed olde . Pa ame e s o da a collec ion Da ase is mainly ocusing on eal-li e affic in he win e condi ions. Desc ip ion o da a collec ion We o ganize he expe imen , by moun ing he senso y amewo k on he ca ’s oo and eco ded da a om en senso s du ing he 90 km ide in he B no, du ing win e in he la e Feb ua y o 2021. In o al, we p o ide abou 3 h o aw da a logs om he ci y cen e, highways, u ban a eas, and coun yside. Da a sou ce loca ion Czechia, B no, 49.195N, 16.608E Da a accessibili y Da a a e a ailable on: h ps://gi hub.com/Robo ics- BUT/B no- U ban- Da ase Rela ed Resea ch A icles B no U ban Da ase (Ligocki e al. 2020) h ps://ieeexplo e.ieee.o g/abs ac /documen /9197277 Fully Au oma ed DCNN-Based The mal Images Anno a ion Using Neu al Ne wo k P e ained on RGB Da a (Ligocki e al. 2021) h ps://www.mdpi.com/1424-8220/21/4/1552 Value o he Da a • The da ase con ains s a e-o - he-a da a om senso s commonly used in obo ics o au- onomous d i ing, like RGB and IR came as, 3D LiDARs, di . RTK GNSS, IMU, o he FMCW ada . All da a a e imes amped. • Da ase ocuses on win e condi ions and includes sessions om he snow-co e ed ci y cen- e , subu b a eas, coun ies, and highways. I is an ideal da a backg ound o de eloping sen- so y usion, obs acle de ec ion, mapping, and localiza ion algo i hms. • The da ase con ains p ecalcula ed objec de ec ion anno a ions by YOLO neu al ne wo k o all came a da a. In o al, he 13 h o de ec ions on 4 RGB and 1 IR came a da a (3.3mil images). I allows p ocessing da a in eal- ime, e en on compu e s wi hou high-pe o mance GPUs. • Da a was al eady used du ing he wo k on scien ific p ojec s [10] . 1. Expe imen al Design, Ma e ials and Me hods 1.1. Senso y equipmen The o iginal se up con ains ou RGB came as wi h 1920 ×1200 px esolu ion (Imaging Sou ce DFK-33GX174 wi h 1/1.2 inch Sony IMX174 chip) and used a he 10 Hz sampling a e. Two came as a e di ec ed o wa d, and each co e s app oxima ely 70 °FoV ( = 8 mm/F1.4 lens). Two la e al came as co e he ca ’s sides wi h 90 °FoV ( = 6 mm/F1.8 lens). All RGB came as combined co e abou 220 °FoV a ound he ca . The RGB ision is enhanced wi h a single nea -in a- ed he mal came a FLIR Tau2 wi h 640 ×512 px esolu ion and 30 Hz ame a e. The came a senses adia ion in a 7.5–13.5 μm spec um ha co esponds wi h a common objec s’ empe a u e in he ange o -40 o 80 °C. This ype o senso p o ides an ad an age in bad ligh condi ions, as i senses he in a- ed adia ions A. Ligocki, A. Jelinek and L. Zalud / Da a in B ie 40 (2022) 107667 3 Fig. 1. (le ) The A las senso y amewo k ins alled on he oo o he es ing ca . Da a we e eco ded in he ea ly Feb ua y o 2021; ( igh ) De ail senso y amewo k o e iew. Fou RGB came as (blue), wo side Velodyne scanne s and single Li ox senso (all g ay), GNSS RTK ecei e (yellow) wi h pai o di e en ial an ennas (whi e), Xsens IMU (o ange) in he cen e , and he FMCW ada ( ed) in he on o he amewo k. ha each objec emi s on i s own, wi hou a need o an ex e nal ligh sou ce. The e o e, i is especially use ul o li ing c ea u es o hea ed ca de ec ion. The came a is equipped wi h a 9 mm lens, ha p o ides 70 °FoV. Nex , we employed wo Velodyne HDL-32 LiDARs wi h 32 lase beams, which sense 20 0 0 poin s pe lase beam pe scan wi h a a e o 10 Hz. In o al senso p oduces abou 0.6 mil poin s pe second. Bo h LiDARs a e sligh ly il ed a ound he o wa d-poin ing axis. Fi s , i gi es a be e co e o a eas nea he ca . Second, as he senso s sense in di e en plains, hey c ea e a g id-like pa e n on objec s isible by bo h senso s and be e co e hem wi h ange mea- su emen s. The ou h senso is he Xsens-MTi-G-710 3D ine ial measu emen uni ha p o ides he 3D linea accele a ion, 3D angula eloci y, 3D magne ic field measu emen , empe a u e, and ai p essu e. Addi ionally, IMU also ecei es low-quali y GNSS da a. Las bu no leas impo an senso in he o iginal se up is he T imble BX 982 RTK GNSS ecei e wi h di e en ial T imble AG25 an ennas enabling heading es ima ion. Fig. 1 shows he da a acquisi ion pla o m and i s ins alla ion on he oo o he ca . Fig. 2 shows he sample da a om he eco ded senso s. Newly, we moun ed a single Li ox Ho izon LiDAR on he on o he senso y amewo k, co - e ing app oxima ely 80 °ho izon al and 25 ° e ical FoV in on o he ca . The senso gene a es abou 240 0 0 0 poin s pe second wi h 2 cm and 0.05 °angula accu acy. Acco ding o he official documen a ion, he LiDAR senses objec s up o 260 m away. In ou case, he objec s can be seen by he LiDAR wi h high ce ain y a a dis ance o 150 m. In con as o he Velodyne HDL-32e, he Li ox Ho izon does no scan he en i e 360 °su - ounding wi h mul iple lase beams, bu i uses a single 905 nm lase ha con inuously co e s he senso ’s on age wi h a non- epe i i e, ose-like pa e n. This se up g ea ly imp o ed poin cloud densi y in he mos c ucial egion in on o he ehicle. The Li ox Ho izon’s da a a e s o ed in he lida _cen e / sub olde o he eco ding in a simila o ma as he Velodyne LiDAR da a. As he Ho izon senso scans he non- epe i i e pa e n and p o ides da a as a con inuous s eam o poin s. We spli his s eam in o 100 ms, ime-o de ed chunks, which a e sa ed as .pcd scans in a common .zip file wi h he co esponding imes amp . x file. Fo ada sensing, we used he TI mmWa e 1642 model. A solid-s a e F equency-Modula ed Con inuous Wa e (FMCW) ada uses an an enna a ay and equency modula ed bu s s o de ec objec s in on o he de ice. Gao e al. [5] desc ibes he measu emen p inciple in de ail. The g ea ad an age o he ada senso is ha he senso p o ides he posi ioning o he de ec ed objec bu using he Dopple ’s e ec also measu es he adial eloci y o he objec wi h espec o he ada . 4 A. Ligocki, A. Jelinek and L. Zalud / Da a in B ie 40 (2022) 107667 Fig. 2. Senso s used o e iew: RGB came as ( op wo ows), IR came a ( hi d ow), and LiDARs (bo om line - le Velodyne, o wa d-looking Li ox, and igh Velodyne scanne s). In ou se up, he ada is configu ed as a 2D senso wi h a scan a e o 30 Hz. I p o ides da a o x, y coo dina es o he de ec ed objec and i s adial eloci y wi h espec o he ada senso . The maximal ange o he de ec ed objec is 50 m, wi h 0.97 m esolu ion, and he max- imal adial eloci y is ±18.5 ms −2 wi h 0.58 ms −2 esolu ion. Also, he ada senses objec in- dependen ly in ligh condi ions, and i wo ks e en du ing he ain o og when he came as a e blinded. Each scan is s o ed in he ada _ i/ sub olde o he eco ding session olde in a . x file. The e is a imes amp o he scan on e e y line and a coun o de ec ed objec s N con inued by 4 N numbe s ep esen ing x, y , and z coo dina es o he de ec ed objec s and hei adial eloci y. The z coo dina e is always ze o in ou case, bu we chose his o ma o u u e compa ibili y. 1.2. YOLO de ec ions We also included p ecalcula ed YOLO 5 [6] objec de ec ions o all RGB came a da a and he IR de ec ions by ou IR-da a- ained YOLO neu al ne wo k [7] o he BUD: WE (B no U ban Da ase : Win e Ex ension). These addi ions we e done no only o he newly eco ded da a bu also e oac i ely o bo h p e iously published sessions published by [1] . Fo gene a ing he RGB objec anno a ions, we used he p e- ained neu al ne wo k on he COCO da ase [8] , co e ing 80 di e en objec classes. Fo au omo i e applica ions, he mos im- po an classes a e pe sons (0), ehicles and bikes (1–7), affic signs (9,11), and li ing c ea u es (14–23). Fo IR images anno a ion, we used he neu al ne wo k ained acco ding o he [7] . The anno a ed classes a e pedes ians (0), bikes (1), ehicles (2), and dogs (16). We made a dedica ed . x file wi h he came a’s name in he /yolo sub olde o e e y eco d- ing session, desc ibing a single objec pe line. Each line con ains in o ma ion abou he de- ec ed objec ’s class, he objec ’s bounding box size and loca ion in he image, and he de ec ion A. Ligocki, A. Jelinek and L. Zalud / Da a in B ie 40 (2022) 107667 5 Fig. 3. Example o neu al ne wo k p ecalcula ed labels o RGB and IR came a da a. In o al, we p o ide anno a ions o app oxima ely 52 h o RGB ideo (13 h o ou came as) a en ps and abou 13 h o anno a ed IR ideo. In o al, i makes abou 3.2. Fig. 4. O e iew o he mos common anno a ed objec s in he RGB and IR domain. confidence. We eco ded all de ec ions wi h confidence g ea e han 0.4 and in e sec ion o e union wi h o he de ec ion g ea e han 0.2. The example o neu al ne wo k de ec ion is shown in Fig. 3 . Fig. 4 shows he dis ibu ion o he anno a ed classes in he image domain. 1.3. Time synch oniza ion To keep all senso s synch onized, we used he GNSS ime ame o all de ices ha ecei e i . Specifically, hese a e he RTK GNSS ecei e , he IMU, and bo h Velodyne LiDARs. The da a cap u e compu e is synch onized di ec ly wi h he GNSS ecei e by he NTP ime se e ha 6 A. Ligocki, A. Jelinek and L. Zalud / Da a in B ie 40 (2022) 107667 uns on he ecei e . To synch onize he RGB came as, we c ea ed a dedica ed mic op ocesso boa d ha ecei es PPS signal and ime da a ia se ial bus om he GNSS ecei e and handles p ecise came a igge ing and imes amp logging. The Li ox Ho izon synch oniza ion is handled by he PTP 2 ime synch oniza ion be ween he senso and he da a-ga he compu e . The only wo senso s ha we could no p ecisely imes amp a e he he mal came a and he ada , as hey do no p o ide an inpu o ou pu igge ing signal wi h he cu en fi mwa e e sion. 1.4. Calib a ion To es ima e he calib a ion pa ame e s o he senso s, we used se e al publicly a ailable ools. Fo RGB came as and IMU pa ame e es ima ion, we used he Kalib ool [2] and [3] o es ima e he ans o ma ion be ween he RGB came as and LiDARs. Fo he mal came a posi ion es ima- ion, we ollow he wo k [4] . The posi ions o GNSS an ennas a e es ima ed based on he ha d- wa e documen a ion and ela ion be ween he an enna’s chassis moun and he ac ual posi ion o he an enna inside he chassis. 1.5. Backend in as uc u e The da a ga he ing sys em’s co e is a ba e y-powe ed desk op compu e wi h AMD Ryzen Th ead ippe 1950x CPU, N idia 1080Ti g aphic ca d, and NMVe SSD disk o high da a band- wid h necessa y o eco ding. The 800 Wh ba e y pack can supply he compu e and he en i e senso y amewo k o abou wo hou s. The RGB and he mal came as, LiDARs, GNSS ecei e , and came a synch oniza ion mic ocon olle boa d a e connec ed ia a D-Link DGS-1510-20 IP swi ch wi h a 10Gb/s line o he compu e . The IMU is connec ed h ough USB, which simula es a i ual se ial bus. 1.6. Reco dings In o al, we ex ended ou exis ing 10 h and 350 km da ase wi h ou new eco ding sessions ha co e app oxima ely 3 h and 90 km o new eal- affic da a in snowy win e condi ions. Th ee sessions a e eco ded du ing dayligh in cloudy wea he . They co e ci y ides as well as high-speed oads, coun yside, o dense u ban locali ies and se e al loop-closing scena ios. The ou h session is eco ded a e he sunse in an ea ly nigh . The eezing ain cached us du ing he las eco ding session ha co e ed senso s and nea ly comple ely blinded all came a senso s in se e al minu es. This session nicely ep esen s ha sh en i onmen al condi ions, which an au oma ed ehicle can ace in p ac ice. The senso s’ eliabili y is s ill an unsol ed p oblem. Fo example, he snow and men ioned eezing ain can be handled wi h he hea ed co e o he senso . On he o he hand, he wa e ha emains on he senso ’s op ics s ill a ec s he mea- su emen . Mo eo e , in condi ions o s ong win e and high eloci y o he ehicle, e en he senso ’s hea ing can ail. This ype o co up ed da a allows de elope s o wo k on algo i hms ha dynamically e alua e he senso s’ eliabili y and adjus he decision-making pipeline. These ou sessions a e di ided in o 13 sho e pa s. See Table 1 o de ails and Fig. 5 depic ing a- e sed ajec o ies. 2. Da a Desc ip ion Each eco ded session pa has a s ic s uc u e and o ma o all eco ded da a. Fo a de- ailed desc ip ion, please see Table 2 . A. Ligocki, A. Jelinek and L. Zalud / Da a in B ie 40 (2022) 107667 7 Table 1 B no u ban da ase : winde ex ension eco ding sessions o e iew. Session Pa En . Dis ance [km] Du a ion [h:mm] Day ime and Wea he 1 1 sub 8,2 00:17 noon - cloudy 2 ci y 13,5 00:30 2 1 coun y 6,8 00:15 noon - cloudy 2 21,3 00:31 3 1 sub 1,4 00:03 a e noon - clody 2 8,9 00:20 3 1,5 00:03 4 4,7 00:11 5 3,3 00:07 4 1 sub 2,8 00:07 nigh - so ain 2 sub 2,4 00:06 3 coun y 3,7 00:07 4 sub 8,7 00:12 To al 13 - 87,2 02:50 - Fig. 5. Visualized ajec o ies a e sed du ing he BUD Win e Ex ension eco ding in B no, Czech Republic. 2.1. Da a s uc u e Each eco ding session appea s in a sepa a e olde , whe e each senso s o es da a in he dedica ed olde . RGB da a a e s o ed as a .mp4 ideos wi h lossless h265 encoding and a co e- sponding . x file ha con ains imes amps o each came a ame. We chose his ideo se up o keep he high quali y o dis ibu ed image da a, and a he same ime o keep he decen size o he da ase . The lossless p ese o he h265 encoding means ha he encoding p ocess bypasses he DCT (disc e e cosine ans o ma ion) and he quan iza ion, bu he p edic ions a e s ill used. The LiDAR poin clouds a e s o ed in a .zip file as a se o .pcd files o each 360 °scan. Again, each scan has he co esponding imes amps in i s accompanying . x file. Fo Li ox LiDAR, each 8 A. Ligocki, A. Jelinek and L. Zalud / Da a in B ie 40 (2022) 107667 Table 2 Da afiles s uc u e and desc ip ion Sub olde File File Desc ip ion; Da a Fo ma calib a ion came a_i .yaml Con ains wid h and heigh o he image and in insic calib a ion pa ame e s, and came a dis o ion pa ame e s came a_le _ on .yaml came a_le _side.yaml came a_ igh _ on .yaml came a_ igh _side.yaml ames.yaml Defines ansla ion and o a ion o each senso wi h espec o he cen e o he ame ( he IMU senso ) came a_i ideo.mp4 512 ×640px@30 ps, h265 ideo imes amp. x Video ames iming and min and max empe a u e on he IR image [ °C]; Line o ma : epoch_ imes amp, ame_no, minimal_ emp, maximal_ emp came a_le _ on ideo.mp4 1920 ×1200px@10 ps, h265 ideo came a_le _side came a_ igh _ on imes amp. x Video ames iming and in e nal came a’s iming; Line o ma : epoch_ imes amp, ame_no, came a_ imes amp came a_ igh _side gnss pose. x WGS84 posi ion; Line o ma : epoch_ imes amp, la i ude, longi ude, al i ude, azimu h ime. x Time o m IMU’s GNSS ecei ee; Line o ma : epoch_ imes amp, yea , mon h, day, hou s, minu es, seconds, nanoseconds imu d_qua . x Senso ’s o ien a ion di e ence om he las sample as a qua e nion; Line o ma : epoch_ imes amp, qua _x, qua _y, qua _z, qua _w gnss. x WGS84 posi ion by IMU’s gnss ecei e ; Line o ma : epoch_ imes amp, la i ude, longi ude, al i ude imu. x 3D linea accele a ion [m/s 2 ], angula eloci y [ ad/s], and IMU’s o ien a ion as qua e nion; Line o ma : epoch_ imes amp, acc_x, acc_y, acc_z, ang_ el_x, ang_ el_y, ang_ el_z, qua _x, qua _y, qua _z, qua _w mag. x 3D magne ic induc ion [G]; Line o ma : epoch_ imes amp, mag_x, mag_y, mag_z p essu e. x A mosphe ic p essu e [Pa]; Line o ma : epoch_ imes amp, p essu e emp. x Senso ’s empe a u e [ °C] Line o ma : epoch_ imes amp, eme a u e ime. x Time om GNSS ecei e ; Line o ma : epoch_ imes amp, yea , mon h, day, hou s, minu es, seconds, nanoseconds lida _cen e scans.zip Zip file ha con ains all .pcd scan files. Each .pcd file con ains a 100 ms da a chunk imes amp. x Timing o LiDAR scans; Line o ma : epoch_ imes amp, scan_numbe lida _le scans.zip Zip file ha con ains all .pcd scan files. Each .pcd file con ains da a om a single ull 360 deg scan by he LiDAR senso lida _ igh imes amp. x Timing o LiDAR scans; Line o ma : epoch_ imes amp, scan_numbe , inne senso ’s imes amp ada _ i scans. x Objec s de ec ed by he FMCW ada , posi ion [m] and adial eloci y [m/s] o each objec ; Line o ma : epoch_ imes amp, no_o _de ec ed_objec , [x_0, y_0, z_0, adial_ el_0, x_1, y_1, ..., adial_ el_n] yolo came a_le _ on . x Objec s de ec ed in he came a da a; Line o ma : ame_numbe , bb_cen e _x, bb_cen e _y, bb_wid h, bb_heigh , de ec ion_confidence, class A. Ligocki, A. Jelinek and L. Zalud / Da a in B ie 40 (2022) 107667 9 .pcd file con ains he se o poin s measu ed du ing he 100 ms ime chunk. Finally, he IMU, ada , and GNSS ecei e s o e da a in ex o ma in sepa a e files, whe e each file co esponds o a specific da a ype (accele a ion, angula eloci y, global pose, e c.). In he case o calib a ion, he e is a file dedica ed o each came a ha con ains i s in insic calib a ion pa ame e s and dis o ion coefficien s. Ano he ex file con ains senso layou in o - ma ion as a ansla ion and o a ion wi h espec o he IMU. The main inspi a ion o chosen da a o ma was he Ox o d Robo Ca Da ase [9] . E hics S a emen The au ho s decla e ha he manusc ip mee s all he ules and condi ions desc ibed in he “E hics in publishing” sec ion ( h ps://www.else ie .com/jou nals/da a- in- b ie /2352- 3409/ guide- o - au ho s ). Du ing he da ase ga he ing, no expe imen s on humans no animals we e in ol ed. All he da a was collec ed in acco dance wi h he law o Czechia. Decla a ion o Compe ing In e es The au ho s decla e ha hey ha e no known compe ing financial in e es s o pe sonal ela- ionships which ha e o could be pe cei ed o ha e influenced he wo k epo ed in his a icle. CRediT Au ho S a emen Adam Ligocki: Visualiza ion, Concep ualiza ion, Da a cu a ion; Ales Jelinek: Visualiza ion, Concep ualiza ion, Da a cu a ion; Ludek Zalud: Supe ision. Acknowledgmen s The wo k has been pe o med in he p ojec A chi ec ECA2030: T us able a chi ec u es wi h accep able esidual isk o he elec ic, connec ed and au oma ed ca s, unde g an ag eemen No 877539 8A20 0 02. The wo k was co- unded by g an s o Minis y o Educa ion, You h and Spo s o he Czech Republic and Elec onic Componen Sys ems o Eu opean Leade ship Join Unde aking (ECSEL JU). The wo k was suppo ed by he in as uc u e o RICAIP ha has e- cei ed unding om he Eu opean Union’s Ho izon 2020 esea ch and inno a ion p og amme unde g an ag eemen No 857306 and om Minis y o Educa ion, You h and Spo s unde OP RDE g an ag eemen No CZ.02.1.01 0.0 0.0 17_043 0010085. The comple ion o his pape was made possible by he g an No. FEKT-S-20-6205 - “Resea ch in Au oma ion, Cybe ne ics, and A ificial In elligence wi hin Indus y 4.0” financially suppo ed by he In e nal science und o B no Uni e si y o Technology. Re e ences [1] A. Ligocki , A. Jelinek , L. Zalud , B no u ban da ase - he new da a o sel -d i ing agen s and mapping asks, in: P oceedings o he IEEE In e na ional Con e ence on Robo ics and Au oma ion (ICRA), IEEE, 2020, pp. 3284–3290 . [2] J. Rehde , J. Nikolic , T. Schneide , T. Hinzmann , R. Siegwa , Ex ending kalib : calib a ing he ex insics o mul i- ple imus and o indi idual axes, in: P oceedings o he IEEE In e na ional Con e ence on Robo ics and Au oma ion (ICRA), IEEE, 2016, pp. 4304–4311 . [3] A. Geige , F. Moosmann , O. Ca , B. Schus e , Au oma ic came a and ange senso calib a ion using a single sho , in: P oceedings o he IEEE In e na ional Con e ence on Robo ics and Au oma ion, IEEE, 2012, pp. 3936–3943 . [4] L. Zalud , P. Kocmano a , Fusion o he he mal imaging and ccd came a based da a o s e eo ision isual elep- esence, in: P oceedings o he IEEE In e na ional Symposium on Sa e y, Secu i y, and Rescue Robo ics (SSRR), IEEE, 2013, pp. 1–6 . [5] X. Gao , G. Xing , S. Roy , H. Liu , Expe imen s wi h mmwa e au omo i e ada es -bed, in: P oceedings o he 53 d Asiloma Con e ence on Signals, Sys ems, and Compu e s, IEEE, 2019, pp. 1–6 .