Describing Behavior Sequences of Fattening Pigs Using Process Mining on Video Data and Automated Pig Behavior Recognition

Ci a ion: Mel sen, A.; Lepsien, A.; Bosselmann, J.; Koschmide , A.; Ha ung, E. Desc ibing Beha io Sequences o Fa ening Pigs Using P ocess Mining on Video Da a and Au oma ed Pig Beha io Recogni ion. Ag icul u e 2023,13, 1639. h ps://doi.o g/10.3390/ ag icul u e13081639 Academic Edi o s: Imke T aulsen and Mehme Gül as Recei ed: 15 July 2023 Re ised: 9 Augus 2023 Accep ed: 16 Augus 2023 Published: 21 Augus 2023 Copy igh : © 2023 by he au ho s. Licensee MDPI, Basel, Swi ze land. This a icle is an open access a icle dis ibu ed unde he e ms and condi ions o he C ea i e Commons A ibu ion (CC BY) license (h ps:// c ea i ecommons.o g/licenses/by/ 4.0/). ag icul u e A icle Desc ibing Beha io Sequences o Fa ening Pigs Using P ocess Mining on Video Da a and Au oma ed Pig Beha io Recogni ion And eas Mel sen 1,* , A id Lepsien 2, Jan Bosselmann 2, Agnes Koschmide 3,4 and Ebe ha d Ha ung 1 1Ins i u e o Ag icul u al Enginee ing, Facul y o Ag icul u al and Nu i ional Sciences, Kiel Uni e si y, 24118 Kiel, Ge many; eha ung@il .uni-kiel.de 2Depa men o Compu e Science, Facul y o Enginee ing, Kiel Uni e si y, 24118 Kiel, Ge many; [email p o ec ed] (A.L.); [email p o ec ed] (J.B.) 3 Business & In o ma ion Sys ems Enginee ing, Facul y o Law, Business and Economics, Uni e si y o Bay eu h, 95447 Bay eu h, Ge many; agnes.koschmide @uni-bay eu h.de 4F aunho e FIT, 95444 Bay eu h, Ge many *Co espondence: amel sen@il .uni-kiel.de Abs ac : This s udy aimed o demons a e he applica ion o p ocess mining on ideo da a o pigs, acili a ing he analysis o beha io al pa e ns. Video da a we e collec ed o e a pe iod o 5 days om a pig pen in a mechanically en ila ed ba n and used o analysis. The app oach in his s udy elies on a se ies o indi idual s eps o allow p ocess mining on his da a se . These s eps include objec de ec ion and acking, spa io empo al ac i i y ecogni ion in ideo da a, and p ocess model analysis. Each s ep gi es insigh s in o pig beha io a di e en ime poin s and loca ions wi hin he pen, o e ing inc easing le els o de ail o desc ibe ypical pig beha io up o p ocess models e lec ing di e en beha io sequences o clus e ed da ase s. Ou da a-d i en app oach p o es sui able o he comp ehensi e analysis o beha io al sequences in con en ional pig a ming. Keywo ds: beha io sequences; p ocess mining; AI ideo analysis; a ening pigs; unc ional a eas 1. In oduc ion P ocess mining is a well-es ablished me hod o gaining insigh in o da a by s uc u ing i in o a sequence o ac i i ies, known as a p ocess model [ 1 ]. The me hod has been success ully applied o a ious domains, including heal hca e, inance, and manu ac u ing and is mainly used o iden i y bo lenecks o compliance issues wi hin p ocesses. Al hough p ocess mining has been p ima ily used in he business con ex , i can also p o ide bene i s o disciplines dealing wi h high olumes and e aci y o da a, such as li e o na u al science. These disciplines, howe e , o en equi e a s uc u ed app oach o answe ing p ocess- ela ed ques ions, like in ou scena io, o iden i y beha io pa e ns o g oups o animals. Obse ing al e a ions in he beha io p ocesses o pigs can be a help ul ool o ana- lyzing and e alua ing animal beha io , animal heal h and en i onmen al impac . Howe e , mos app oaches on iden i ying pig beha io based on ideo da a so a mos ly ocus on single speci ic ac i i ies e.g., eeding/d inking ecogni ion, ail bi ing, playing beha io , o agg ession ecogni ion [2]. Di e en lying pa e ns o pigs can be iden i ied wi h compu e ision-based moni- o ing o gi e an indica ion o animal heal h, wel a e and he mal com o s a e o g oup- housed pigs indica ing clima e condi ions in mechanically en ila ed ba ns. Likewise obse a ions o ac i i y and eed in ake, which a y depending on di e en clima e condi- ions, suppo he con ol o he abo e [3]. Analyzing he eeding beha io o pigs suppo s he indica ion o heal h issues and can play an impo an ole in he b eeding p ocess. Me hods o iden i ying eeding beha io based on ideo da a ha e been published wi h a good success a e [4]. Ag icul u e 2023,13, 1639. h ps://doi.o g/10.3390/ag icul u e13081639 h ps://www.mdpi.com/jou nal/ag icul u e Ag icul u e 2023,13, 1639 2 o 20 Beha io changes, like slowdown and weakening, can be used as an ea ly wa ning indi- ca o o pa hological in ec ions. De ec ing hese beha io al changes as pa o su eillance in eal- ime suppo s a ms in hei moni o ing asks [ 5 ]. S udies o Be gamini e al., 2021 [6] ha e shown ha gene al beha io al changes can also be ex ac ed om la ge compu e - ision da ase s and ma k long- e m changes in i e undamen al indi idual beha io pa e ns. Those e ec s on gene al beha io such as lying, mo ing, eeding, d inking and elimina ing, a e also ele an in e ms o , en i onmen al impac . Pigs a e known o ha e he abili y o s uc u e hei pens in o unc ional a eas o elimina ion, eeding, and sleeping i ce ain condi ions in e ms o , e.g., space allowance a e me . This beha io is in luenced by hei inna e need o hygiene and com o , as well as hei social hie a chies and pa e ns o use. S udies ha e shown ha pigs end o es ablish clea bounda ies be ween elimina ion and sleeping a eas, wi h he o me being loca ed a a dis ance om he la e [ 7 ]. Feed- ing a eas, on he o he hand, a e o en loca ed nea bo h elimina ion and sleeping a eas, e lec ing hei impo ance in he daily ou ine o pigs [ 8 ]. Sel -s uc u ed pig pens consis o a de ined soiling a ea in which no mal elimina o y beha io akes place [ 7 ]. In case o de ia ing elimina ion beha io due o changes in pig densi y, pa i ion ype o changes in clima e condi ions [ 8 , 9 ] po en ially inc eased ammonia emissions ha will nega i ely in luence he en i onmen al impac ela ed o he size o he soiled a ea co esponds o he eleased ammonia emissions [10]. Also, he ea ly de ec ion o heal h and wel a e comp omises such as he iden i ica ion o clinical and subclinical illnesses can be obse ed based on ecognized gene al beha io al changes in pigs wi h esul s om ideo da a [ 11 , 12 ]. Simila epo s exis o he analysis o social in e ac ions such as agonis ic beha io a e eg ouping as an ea ly indica o o misma ches wi hin he g oup [13–15]. This pape p oposes o use p ocess mining as a me hod o iden i y beha io al pa e ns o a ening pigs and he ela ed p ocesses om ideo da a. The goal is o demons a e ha p ocess mining can be a aluable ool o unde s anding he beha io o pigs in hei beha io sequences and gaining addi ional in o ma ion in a empo al, spa ial esolu ion abou he di ision o he pen in unc ional a eas compa ed o he s a e-o - he a me hod. This publica ion explains he s ep-by-s ep app oach leading o he e alua ion h ough he me hod o p ocess mining. I sys ema ically demons a es he p ocedu e by which he analysis un olds. The indi idual s ages include: (1) iden i ica ion o p ocess ac i i ies, (2) iden i ica ion o dis inc beha io al pa e ns and assessmen o hei accu acy, (3) depic- ion o beha io al pa e ns in empo al and spa ial esolu ion, and (4) ans o ma ion o beha io al pa e ns in o p ocess models. The me hodology in his pape is gene ally based on he app oach p esen ed in [ 16 ] and ex ends i in e ms o adap ing he me hodology o a eal-li e use case, an imp o ed e alua ion, and an in-dep h discussion o he analysis esul s. In [ 16 ], he iden ical ideo da ase was used o e alua e he p oposed me hods, wi h b ie analysis esul s ela ed o hei capabili y o p oduce meaning ul ou comes. This pape demons a es he adap ed app oach o analyzing ideo eco dings o a ening pigs in subsequen s eps o he da a- d i en app oach o ex ac aluable in o ma ion abou hei o e all beha io and beha io sequences. The da a-d i en app oach’s esul s can aid in imp o ing decision-making ega ding pig beha io sequences. 2. Ma e ials and Me hods 2.1. Animals and Housing Video eco dings o a ening pigs we e eco ded a he Teaching and Resea ch Facili y Fu e kamp o he Schleswig-Hols ein Chambe o Ag icul u e. In o al, 1400 Pigs we e housed in 14 compa men s in he expe imen al ba n on si e. Each compa men was mechanically en ila ed and subdi ided in o en pens ( i e pe side—one behind ano he , di ided by an inspec ion walkway) each 4.08 m wide × 2.74 m long, each consis ing o 11 pigs. Thus, each pig had 1.02 m2o space a i s disposal. Ag icul u e 2023,13, 1639 3 o 20 All a ening pigs we e ea ed in con en ional pig pens wi h mechanical en ila ion. The space allowance in his in es iga ion was clea ly abo e he minimum se by he Eu opean o Ge man legisla ion based on li e weigh [17,18]. Insu icien space can cause nega i e social beha io owa d penma es, esul ing in skin lesions, lameness, ail bi ing, o educed g ow h [ 19 ]. In addi ion o he space allowance he beha io o a ening pigs can be in luenced by a ious ac o s, such as diu nal hy hms, eeding, social in e ac ions, and en i onmen al condi ions. All condi ions we e kep unchanged du ing he es ing pe iod o minimize ex e nal in e e ences. The ba n clima e in he s udy pe iod was no in luenced by he s udy, bu ollowed he pa ame e s o he s udy ba n used in p ac ice. The whi e luo escen ube ligh s we e swi ched on du ing he day ime (06:00–16:00 o’clock). All pens we e equipped wi h ully-sla ed conc e e loo s and an unde - loo slu y pi . Pigs we e ed up o se en imes pe day ( eeding imes: 07:33, 09:25, 10:50, 12:31, 14:52, 16:52 on all days and addi ionally a 17:44 on he i s wo days) wi h a p ede ined amoun o liquid eed in a long ough sys em, gi ing enough eeding space o all pigs o simul aneously ha e access o he ough. In gene al, he eeding a io was adjus ed in h ee phases acco ding o he weigh o he animals (ea ly 25 kg, middle 40 kg, la e 70 kg). Du ing he in es iga ions all pigs we e ed wi h he i s eed. As en ichmen ma e ial sisal opes we e a ached o one wall o he pen. A ske ch o he pen design wi h de ails like posi ion o eeding ough as well as he posi ion o he came as wi hin he compa men can be seen in Figu e 1. Ag icul u e 2023, 13, x 3 o 21 mechanically en ila ed and subdi ided in o en pens ( i e pe side—one behind ano he , di ided by an inspec ion walkway) each 4.08 m wide × 2.74 m long, each consis ing o 11 pigs. Thus, each pig had 1.02 m² o space a i s disposal. All a ening pigs we e ea ed in con en ional pig pens wi h mechanical en ila ion. The space allowance in his in es iga ion was clea ly abo e he minimum se by he Eu- opean o Ge man legisla ion based on li e weigh [17,18]. Insufficien space can cause nega i e social beha io owa d penma es, esul ing in skin lesions, lameness, ail bi ing, o educed g ow h [19]. In addi ion o he space allowance he beha io o a ening pigs can be in luenced by a ious ac o s, such as diu nal hy hms, eeding, social in e ac ions, and en i onmen al condi ions. All condi ions we e kep unchanged du ing he es ing pe- iod o minimize ex e nal in e e ences. The ba n clima e in he s udy pe iod was no in luenced by he s udy, bu ollowed he pa ame e s o he s udy ba n used in p ac ice. The whi e luo escen ube ligh s we e swi ched on du ing he day ime (06:00–16:00 o’clock). All pens we e equipped wi h ully-sla ed conc e e loo s and an unde - loo slu y pi . Pigs we e ed up o se en imes pe day ( eeding imes: 07:33, 09:25, 10:50, 12:31, 14:52, 16:52 on all days and addi ionally a 17:44 on he i s wo days) wi h a p ede ined amoun o liquid eed in a long ough sys em, gi ing enough eeding space o all pigs o simul aneously ha e access o he ough. In gene al, he eeding a io was adjus ed in h ee phases acco ding o he weigh o he animals (ea ly 25 kg, middle 40 kg, la e 70 kg). Du ing he in es iga ions all pigs we e ed wi h he i s eed. As en ichmen ma e ial sisal opes we e a ached o one wall o he pen. A ske ch o he pen design wi h de ails like posi ion o eeding ough as well as he posi ion o he came as wi hin he compa men can be seen in Figu e 1. Figu e 1. Floo plan o compa men 6 wi h he came a (A) acing pen 68 which was used o ideo eco ding. (B) Posi ion o he eeding ough. The 11 boa s o Topigs, PIC and DK gene ic in pen #68 we e on a e age 107.6 d (SD 4.3 d) and weighed 56.06 kg (SD 5.25 kg). All pigs we e andomly dis ibu ed e enly among he compa men s acco ding o gene ics and weigh . Figu e 1. Floo plan o compa men 6 wi h he came a (A) acing pen 68 which was used o ideo eco ding. (B) Posi ion o he eeding ough. The 11 boa s o Topigs, PIC and DK gene ic in pen #68 we e on a e age 107.6 d (SD 4.3 d) and weighed 56.06 kg (SD 5.25 kg). All pigs we e andomly dis ibu ed e enly among he compa men s acco ding o gene ics and weigh . 2.2. Video Reco dings To moni o he beha io o pigs, a ideo eco ding se up was used consis ing o a came a and an HDD ecei e uni ANNKE H500 (ANNKE Secu i y Technology Inc., Rowland Heigh s, CA, USA). The came a was posi ioned acing he pigs This placemen Ag icul u e 2023,13, 1639 4 o 20 allowed o a comp ehensi e iew o almos he en i e pigpen and cap u ed he mo emen s o he pigs. Due o he low ceiling le el he came as we e posi ioned a a heigh o app oxima ely 2.9 m abo e he loo and angled downwa d a 45 ◦ o ensu e ha almos he en i e pen was cap u ed in he ield o iew. The came as we e se o a esolu ion o 1080 p and a ame a e o 12.5 ames pe second. This se ing allowed o clea and de ailed eco dings o he pigs’ beha io . The came as we e also se o in a ed mode o cap u e he eco dings in low-ligh condi ions which was he case as soon as he ligh s we e u ned o . All ideo eco dings we e s o ed in he HDD ecei e uni in a o ma ha was compa ible wi h he so wa e used o da a analysis. The ideo eco dings we e aken in he ime pe iod om 12 No embe 2021 o 10 Decembe 2021 be ween 06:00 am and 06:00 pm. This ime pe iod was chosen because i co e ed he majo i y o he pigs’ ac i e hou s as ac i i y le el signi ican ly dec eases du ing nigh hou s [ 20 ]. The eco dings we e aken con inuously wi hou in e up ion o ensu e ha no beha io was missed. 2.3. P ep ocessing o Video Da a Fi s , he ideo eco dings we e p epa ed by (1) selec ing he ideo segmen s ele an o he analysis being conduc ed and (2) o ganizing he selec ed ideo segmen s in o a s uc u ed da ase acco ding o a uni o m o ma . Since he pu pose o he analysis does no impose any es ic ions on he ime o day o loca ion o he beha io s o be moni o ed, all eco dings we e in p inciple ele an . Howe e , inal eco dings om i e consecu i e days (13 No embe 2021–17 No embe 2021) we e selec ed and all o he eco dings we e excluded om he analysis, as p ocessing esou ces we e limi ed and hese selec ed days we e su icien o cap u e ep esen a i e beha io s. The selec ed ideo eco dings, which we e spli in o mul iple ideo iles pe day due o he eco ding se up, we e consolida ed in o a single ile pe selec ed day. To educe p ocessing ime, he 1080 px esolu ion o he ideo da a was educed o 854 × 480 px. While downscaling ideo da a signi ican ly educes p ocessing ime o ideo analysis, excessi e downscaling can esul in loss o in o ma ion (e.g., due o small objec s becoming blu ed in he images). Fo he analysis in his s udy, small de ails we e negligible, because he analysis o b oad beha io al classes did no equi e ine-g ained de ails o dis inguish di e en a ian s o he same class, and he e o e he downscaling o 854 × 480 px p omised o be an app op ia e solu ion. Since he models used la e o ac i i y ecogni ion equi ed a ame a e o 30 ps as inpu , ideos we e esampled o ha ame a e using FFmpeg (FFmpeg p ojec ). The eco dings also cap u ed some beha io in neighbo ing pig pens. To educe noise in p ocessing and he analysis, a s a ic mask was o e laid o e he ideo iles, so ha only he a ge ed pig pen was isible. To quan i y he amoun o ac i i y isible in a ideo sequence wi hou ha ing o ac ually de ec he mo emen and execu e ac i i ies o he pigs, a cus om ac i i y sco e based on pixel changes be ween ames we e used [ 21 ]. To calcula e he ac i i y sco e du ing a gi en ideo, ames a e sampled a a ixed a e and compa ed o he p e iously sampled ame. Each ame was di ided in o a 20 × 20 g id and he mean alues o each g id we e calcula ed a a ixed sampling a e o one ame e e y 30 s. Due o he di e en ligh ing condi ions du ing he day, each ame was con e ed o a g ayscale image be o e pe o ming he calcula ions. The ac i i y sco e was hen calcula ed as he sum o he absolu e di e ences o he pixel alues o each co esponding g id be ween wo sampled ames. The alues we e smoo hed using a mo ing a e age wi h a window size o 60 samples, which means ha he inal ac i i y sco e a a gi en poin in ime akes in o accoun alues om i s pas and u u e 15 min. Depending on he goal o he analysis, he ac i i y sco e could be used o il e ou i ele an segmen s o he ideo eco dings (e.g., segmen s whe e gene ally a low amoun o ac i i y is obse ed). Ag icul u e 2023,13, 1639 5 o 20 2.4. De ec ion and T acking The nex s ep o he analysis in ol ed de ec ing he posi ions o pigs in he ideo and acking pigs h oughou he ideo eco dings. Fo pig de ec ion, a da ase o 614 anno a ed images was c ea ed by andomly sampling ames om he ull se o ideo eco dings and manually labeling he pigs isible in hese ames wi h bounding boxes. This da a se was andomly spli in o aining and alida ion se s a an 80:20 spli . The aining da a se was used o ine une a YOLO 7 classi ie [ 22 ], which was p e- ained on ImageNe [ 23 ]. By using he pa ame e s o a ne wo k p e ained on ImageNe and ine- uning o he cus om ask (in ou case he pig de ec ion), he amoun o equi ed aining images was signi ican ly educed compa ed o aining om sc a ch [24]. Fo mul iple objec acking, a By eT ack acke [ 25 ] was used. By eT ack is a acking- by-de ec ion me hod, i.e., i uses de ec ions p o ided by a sepa a e objec de ec o and co ela es he de ec ions ac oss ames h ough in e nal mo ion models. This p o ides he lexibili y o easily adap he acke o new se ings and use cases, which can be implemen ed by simply adap ing he objec de ec o o he new se ing (e.g., a di e en came a pe spec i e, a chi ec u e, o species) [26]. 2.5. Ac i i y Recogni ion (Labeling o Animal Beha io ) The nex s ep a e localizing and acking he pigs is he de ec ion o hei ac i i ies. To suppo he ans e abili y o ou app oach o o he se ings and analysis goals, a gene ic ac i i y ecogni ion was equi ed allowing o de ec mul iple concu en beha io s occu ing in a ideo. Mos exis ing me hods, howe e , o pig beha io ecogni ion ei he allow he de ec ion o single ac i i ies o speci ic g oups o ac i i ies [ 2 ]. O hose solu ions ha a e able o de ec mul iple di e en ac i i y classes, mos s ill use ac i i y-speci ic assump ions and algo i hms (e.g., a pig wi h i s head posi ioned a a known posi ion o a d inke is classi ied as d inking in [ 6 ]). To educe he need o explici ly impose any assump ions in o he ac i i y ecogni ion, we selec ed a deep lea ning me hod based on CNNs (Con olu ional Neu al Ne wo ks) ha lea ns o de ec ac i i ies based on labeled aining examples o ac i i ies occu ing in a ideo. In pa icula , SlowFas [ 27 ] was used o spa io- empo al ac ion de ec ion, which is a ask om compu e ision conce ned wi h de ec ing he ac i i ies concu en ly pe o med by mul iple objec s in a ideo. Li e al., 2020 [ 28 ] showed ha spa io- empo al CNNs can success ully be applied o pig beha io ecogni ion. They speci ically op imized he design o a SlowFas -based CNN o pig beha io ecogni ion. In hei e alua ion, his design signi ican ly imp o ed accu acy o unseen se ings (e.g., a came a pe spec i e di e en om hose obse ed in he aining se ), bu he accu acy was compa able o mo e gene ic ne wo k designs (i he se ing is simila in aining and in e ence). Gene ally, ne wo ks p e- ained on Kine ics [ 29 ], which is a ideo da ase o ac i i ies pe o med by humans, pe o med be e han he non-p e- ained ne wo ks, wi h he excep ion o he ne wo k specially designed o pig beha io ecogni ion. As o he analysis in his s udy, he se ing o he aining and analysis ideos was iden ical, we chose o use a gene ic SlowFas 4 × 16 ne wo k p e- ained on Kine ics. A cus om aining da ase was p epa ed by sampling sho ideo sequences om he ull se o ideo eco dings, con aining bo h RGB and in a ed segmen s as well as segmen s wi h high and low gene al ac i i y. The beha io obse ed in hese aining sequences was manually anno a ed by a ained obse e acco ding o he beha io classes de ined in Table 1, esul ing in a o al o 9240 anno a ed samples o ac i i ies. This da ase was di ided in o 70:30 aining/ alida ion spli s, and used o ine- une he p e- ained SlowFas 4×16 model. P e- aining using he da ase p o ided by Be gamini e al., 2021 [ 6 ] was also e alua ed, bu did no yield any signi ican imp o emen s in aining con e gence speed o quali y. Ag icul u e 2023,13, 1639 6 o 20 Table 1. Beha io s o pigs and he espec i e de ini ions as used in he da ase . Beha io De ini ion lying Pig in a es ing posi ion, ypically lying down on he side o he unk o ches /s e num, wi h minimal mo emen and i s body suppo ed by he loo . si ing Con ac o he loo wi h he ee o he on legs and he pos e io po ion o he pig’s body. s anding Con ac o he loo wi h all ee wi hou changing posi ion. mo ing Pig in mo ion, displaying walking o unning beha io , wi h all ou legs ac i ely mo ing hei body o wa d. in es iga ing Pig using i s snou o sea ch and dig in o he g ound o o he su aces, o en in a epe i i e manne , as pa o i s na u al beha io o ind ood o explo e he su oundings. eeding Focused engagemen wi h a ood sou ce, cha ac e ized by epea ed chewing, swallowing, and oo ing beha io . de eca ing Ac o elimina ion, seen as a s anding posi ion, ollowed by he expulsion o eces om he body. playing Ac i e engagemen wi h occupa ion ma e ial o pigs like opes. miscellaneous All o he ac i i ies ha canno be alloca ed o any o he abo e ac i i ies. The ained model was hen used o de ec he ac i i ies execu ed by he pigs in he comple e i e days o ideo eco dings selec ed o analysis. The esul s o ac i i y ecogni ion indica e he loca ion and ime o each de ec ed ac i i y, and can be used in combina ion wi h he p e iously ex ac ed acking in o ma ion o econs uc sequences o ac i i ies pe o med by speci ic pigs. 2.6. P ocess Mining The inal pa o he analysis is he applica ion o p ocess mining o he ex ac ed ac i i ies. P ocess mining algo i hms ypically equi e s uc u ed e en da a a a high le el o abs ac ion, i.e., a log o e en s desc ibing when each ac o s a s and ends wi h he execu ion o speci ic ac i i ies, which in u n e e s o speci ic ins ances o a p ocess (cases). Albei al eady e e ing o he abs ac ac i i ies ele an o he conduc ed p ocess mining analysis, he esul s om ac i i y ecogni ion a e a a much lowe le el o abs ac ion, because hey simply lis he ac i i ies de ec ed o each pig a egula ime in e als, and a e no o ganized in o cases. To ans o m he ac i i y ecogni ion esul s in o a s uc u e complying wi h p ocess mining, he me hod p esen ed in [ 16 ] was used. Speci ically, E en Abs ac ion and Case Co ela ion we e used o p epa e he inpu o he p ocess mining algo i hms. Fo E en Abs ac ion, mul iple subsequen ly ollowed de ec ions o he same ac i i y o he same pig a e agg ega ed o a single ins ance o his ac i i y using he empo al agg ega ion echnique wi h smoo hing om [ 16 ]. In ypical p ocess mining applica ions, whe e he unde lying p ocess is a well-s uc u ed business p ocess, cases can ypically be de ined by he s a - and endpoin s o he p ocess execu ions (e.g., a business p ocess could s a when an o de o a p oduc is ecei ed and inish when he p oduc has been shipped). In he daily beha io o pigs, howe e , no such na u al no ion o a case exis s. To cons uc cases in a way ha each case ep esen s an ins ance o he same p ocess, Case Co ela ion is pe o med using de ined s a and end ac i i ies using he echnique om [ 16 ]. Fo his, lying was de ined as bo h he s a and end ac i i y o each p ocess ins ance. This means each ins ance o he analyzed p ocess s a s when a pig s ands up and ends when his same pig lies back down, and he e o e cap u es “ac i e” phases o he pigs. As he beha io o pigs is uns uc u ed and chao ic, ace clus e ing [ 30 ] is hen applied o he e en da a. T ace clus e ing me hods di ide he e en da a o an uns uc u ed p ocess in o mul iple, sepa a e clus e s e e ing o simila beha io . Wi h ace clus e ing, p ocess mining me hods a e applied o each clus e sepa a ely, which ypically yields mo e Ag icul u e 2023,13, 1639 7 o 20 s uc u ed esul s. Fo ou analysis, ea u es such as he occu ence and equency o e en s, and he di ec ly- ollows ela ionships (i.e., i mo ing occu s di ec ly a e lying in a case, mo ing di ec ly ollows lying) in a case we e ex ac ed, scaled by i s a ge ing a mean a ound ze o and hen scaling o uni a iance, educed wi h PCA se o 99% a iance, and o ganized in o 15 clus e s using he k-means algo i hm. Finally, he clus e ed e en da a we e expo ed in o a ep esen a ion p ocessed by p ocess mining ools and p ocess models we e disco e ed o each o he clus e s wi h he p ocess mining so wa e Disco (Fluxicon BV, Eindho en, The Ne he lands). 3. Resul s The esul s co e he quan i ica ion o ac i i ies in spa ial and empo al esolu ion wi h a ocus on he de elopmen o a p ocess mining analysis o beha io ecogni ion in pigs. Fo p ocess mining analysis, i s a pu pose needs o be de ined [ 31 ]. This pu pose guides he analysis wi h espec o how he da a a e p epa ed, how e en s a e ex ac ed and p ocessed, and which p ocess mining algo i hms a e applied o he ex ac ed e en da a. Fo he analysis in his s udy, he daily beha io o pigs was moni o ed wi h he pu pose o ex ac ing pa e ns desc ibing he beha io o pigs in ac i e phases, wi h he goal o enabling au oma ic moni o ing o he ac i i y o pigs. The s epwise app oach including gene al ac i i y quan i ica ion, iden i ica ion o dis inc beha io al pa e ns, and inally ans o ming he beha io al pa e ns in o p ocess models, was pe o med o achie e his goal. The esul s o hese s eps a e p esen ed in he ollowing chap e . 3.1. Gene al Ac i i y in he Pen—Quan i ica ion o Ac i i ies in Spa ial and Tempo al Resolu ion Figu e 2shows a e age ac i i y sco es o he whole pen o e ime du ing eco ding imes (06:00–18:00) o di e en obse a ion days. Clea ly ecognized ac i i y pa e ns appea a speci ic imes du ing he day ha co espond wi h eeding imes and he gene al ba n ou ine. Fixed ime pe iods can be iden i ied acco ding o an inc ease in ac i i y occu ence, e.g., simila inc eases in ac i i y a all obse a ion days a e no ed a ound 07:30–08:15, 09:15–10:15, 10:45–11:30, 12:45–13:30, 14:45–15:30, 16:45–17:15 and addi ionally a 17:45 un il obse a ion end. Fu he mo e, a link can be se o he eeding which akes place oge he wi h he con ol walk be ween 07:10 and 07:30. Ag icul u e 2023, 13, x 8 o 21 Figu e 2. Ac i i y sco es o e ime du ing eco ding imes (06:00–18:00) o i e consecu i e obse - a ion days (13 No embe 2021–17 No embe 2021). Ac i i y pa e ns du ing he day may also be in luenced by a i icial ligh ing in he ba n. Ligh ing was ac i a ed om 06:00 un il he las con ol walk (be ween 15:00 and 16:00). Ou side o hese imes, an o ien a ion ligh was used o he animals. In addi ion o he mo e seman ically abs ac ac i i ies, uni o m pe iods o es we e obse ed be- ween he i s obse a ion isi o he s aff and he i s eeding and be ween he i s wo eeding e en s espec i ely, du ing which he ac i i y sco e was signi ican ly educed on all days. In con as , ac i i ies a e he hi d and ou h eeding phases showed a highly he e ogeneous dis ibu ion. Howe e , no signi ican conclusion can be made abou he speci ic ac i i ies (e.g., eeding, lying, de eca ing, e c.) o whe e hese ac i i ies ook place. Figu e 3 illus a es he spa ial dis ibu ion o ela i e ac i i y sco es wi hin he pen a diffe en ime poin s o one speci ic day (13 No embe 2021). Diffe en ac i i y pa e ns can be obse ed be o e he i s eeding phase in he mo ning be ween 06:15–06:45, du ing he second eeding phase be ween 09:15–09:45 and du ing a he e ogeneous ac i i y phase be ween 15:30–16:00. The spa ial dis ibu ion o he ela i e ac i i y sco e shows clea diffe ences o ac i - i y oci wi hin he bay a diffe en imes o he day, wi h a e y low numbe o ac i i ies in he ea ly mo ning hou s be o e he i s eeding and a highe numbe o ac i i ies a eeding imes, as well as in he la e a e noon hou s be ween indi idual eeding pe iods. Figu e 2. Ac i i y sco es o e ime du ing eco ding imes (06:00–18:00) o i e consecu i e obse a- ion days (13 No embe 2021–17 No embe 2021). Ag icul u e 2023,13, 1639 8 o 20 Ac i i y pa e ns du ing he day may also be in luenced by a i icial ligh ing in he ba n. Ligh ing was ac i a ed om 06:00 un il he las con ol walk (be ween 15:00 and 16:00). Ou side o hese imes, an o ien a ion ligh was used o he animals. In addi ion o he mo e seman ically abs ac ac i i ies, uni o m pe iods o es we e obse ed be ween he i s obse a ion isi o he s a and he i s eeding and be ween he i s wo eeding e en s espec i ely, du ing which he ac i i y sco e was signi ican ly educed on all days. In con as , ac i i ies a e he hi d and ou h eeding phases showed a highly he e ogeneous dis ibu ion. Howe e , no signi ican conclusion can be made abou he speci ic ac i i ies (e.g., eeding, lying, de eca ing, e c.) o whe e hese ac i i ies ook place. Figu e 3illus a es he spa ial dis ibu ion o ela i e ac i i y sco es wi hin he pen a di e en ime poin s o one speci ic day (13 No embe 2021). Di e en ac i i y pa e ns can be obse ed be o e he i s eeding phase in he mo ning be ween 06:15–06:45, du ing he second eeding phase be ween 09:15–09:45 and du ing a he e ogeneous ac i i y phase be ween 15:30–16:00. Ag icul u e 2023, 13, x 9 o 21 Figu e 3. Spa ial dis ibu ion o ela i e ac i i y sco es wi hin he pen a diffe en ime poin s on one speci ic day (13 No embe 2021). Du ing he second eeding phase, ac i i y inc eased in he a ea o he eeding ough, which indica es i s link wi h eeding beha io bu also in he lying a ea o he pen. He e - ogeneous dis ibu ions o ac i i y (ac i i y sco e), dis ibu ed h oughou he pen wi h no ocal poin o ac i i y, can be obse ed in he hea map (shown in Figu e 3) om he pe iod o he la e a e noon, which also shows highe ac i i y in he ecal a ea in he op igh co ne . Howe e , he ac i i y sco e is no sufficien o conclude wha ype o beha io is being unde aken in each a ea o he pen. 3.2. Accu acy Objec , T acking and Beha io Recogni ion The objec de ec o was e alua ed using he alida ion spli o he pig de ec ion da- ase , esul ing in a [email protected]:0.95 o 0.864. In he selec ed ideos, on a e age 10.6 o 11 pigs (wi h a s anda d de ia ion o 0.64) we e de ec ed pe ame. The main challenge o he objec de ec o was isual occlusion, i.e., when a pig was hidden om he iew o he came a by o he pigs’ bodies. When mo e han 11 pigs we e de ec ed in a ame, he bounding boxes wi h he lowes de ec ion con idence we e disca ded such ha 11 de ec- ions emained. To e alua e objec acking pe o mance, an e alua ion da ase o 14 1-minu e ideo sequences was sampled om he eco dings and manually anno a ed wi h g ound- u h acking in o ma ion. The acking pe o mance me ics calcula ed by compa ing he acking esul s on hese ideos o he g ound- u h anno a ions a e epo ed in Table 2. When applied o he selec ed ideos, he ackle s eached an a e age leng h o 18 min be o e an ID swi ch o ackle agmen a ion, and he longes ackle s spanned o e mul- iple hou s. While his is no op imal (which would p oduce one 12-h ackle pe pig pe day), he a e age ackle leng h was sufficien o cap u e beha io al pa e ns consis ing o mul iple, sequen ially execu ed ac i i ies. Figu e 3. Spa ial dis ibu ion o ela i e ac i i y sco es wi hin he pen a di e en ime poin s on one speci ic day (13 No embe 2021). The spa ial dis ibu ion o he ela i e ac i i y sco e shows clea di e ences o ac i i y oci wi hin he bay a di e en imes o he day, wi h a e y low numbe o ac i i ies in he ea ly mo ning hou s be o e he i s eeding and a highe numbe o ac i i ies a eeding imes, as well as in he la e a e noon hou s be ween indi idual eeding pe iods. Du ing he second eeding phase, ac i i y inc eased in he a ea o he eeding ough, which indica es i s link wi h eeding beha io bu also in he lying a ea o he pen. He e o- geneous dis ibu ions o ac i i y (ac i i y sco e), dis ibu ed h oughou he pen wi h no ocal poin o ac i i y, can be obse ed in he hea map (shown in Figu e 3) om he pe iod o he la e a e noon, which also shows highe ac i i y in he ecal a ea in he op igh co ne . Howe e , he ac i i y sco e is no su icien o conclude wha ype o beha io is being unde aken in each a ea o he pen. Ag icul u e 2023,13, 1639 9 o 20 3.2. Accu acy Objec , T acking and Beha io Recogni ion The objec de ec o was e alua ed using he alida ion spli o he pig de ec ion da ase , esul ing in a [email p o ec ed]:0.95 o 0.864. In he selec ed ideos, on a e age 10.6 o 11 pigs (wi h a s anda d de ia ion o 0.64) we e de ec ed pe ame. The main challenge o he objec de ec o was isual occlusion, i.e., when a pig was hidden om he iew o he came a by o he pigs’ bodies. When mo e han 11 pigs we e de ec ed in a ame, he bounding boxes wi h he lowes de ec ion con idence we e disca ded such ha 11 de ec ions emained. To e alua e objec acking pe o mance, an e alua ion da ase o 14 1-minu e ideo sequences was sampled om he eco dings and manually anno a ed wi h g ound- u h acking in o ma ion. The acking pe o mance me ics calcula ed by compa ing he acking esul s on hese ideos o he g ound- u h anno a ions a e epo ed in Table 2. When applied o he selec ed ideos, he ackle s eached an a e age leng h o 18 min be o e an ID swi ch o ackle agmen a ion, and he longes ackle s spanned o e mul iple hou s. While his is no op imal (which would p oduce one 12-h ackle pe pig pe day), he a e age ackle leng h was su icien o cap u e beha io al pa e ns consis ing o mul iple, sequen ially execu ed ac i i ies. Table 2. Calcula ed acking pe o mance me ics a e aged o e all e alua ion sequences (MOTA: mul iple objec acking accu acy, IDF1: global min-cos F1 sco e o ID associa ions, # Swi ches: o al numbe o ack swi ches, # F agmen a ions: o al numbe o swi ches om acked o no acked). MOTA IDF1 # Swi ches # F agmen a ions A e age 0.982 0.985 0.5 1.286 S anda d De ia ion 0.032 0.017 0.76 1.541 A [email p o ec ed] (mean a e age p ecision) o 0.7365 was achie ed on he ac i i y ecogni ion alida ion se by he ained ac i i y ecogni ion model. The a e age p ecision o each ac i i y class is lis ed in Table 3. Si ing and s anding showed he lowes a e age p ecision alues by a la ge ma gin. Visual inspec ion o esul s has shown ha hese ac ions a e occasionally con used wi h lying by he model, due o simila i y in isual appea ance and lack o mo emen . Good a e age p ecision alues we e achie ed o common ac i i ies like lying (0.997), eeding (0.994), de eca ing (0.960) and playing (0.969). Table 3. A e age p ecision pe ac i i y class as obse ed in model alida ion. Ac i i y [email p o ec ed] lying 0.997 si ing 0.374 s anding 0.296 mo ing 0.807 in es iga ing 0.793 eeding 0.994 de eca ing 0.960 playing 0.969 miscellaneous 0.439 3.3. Beha io Pa e ns and U iliza ion o he Pen All he a o emen ioned beha io s we e acked h oughou he en i e obse a ion pe iod. Video da a based beha io ecogni ion esul ed in p edic ions o di e en beha io pa e ns. In Table 4, he desc ip i e s a is ics o beha io g ouped in o 10-min blocks can be obse ed. These 10-min blocks p o ided a quan i ying o e iew o he di e se beha io al pa e ns obse ed. The use o hese 10-min blocks allowed o a su icien ly de ailed Ag icul u e 2023,13, 1639 16 o 20 4.2.3. Lying/Res ing Compa able o Ekkel e al., (2003) and Ruckebush e al., (1972) [ 38 , 45 ], he indings in his s udy highligh lying beha io as he dominan ac i i y, accoun ing o up o 80% o he o al ime budge , while he pe cen age o lying dec eased du ing he la e hal o he day. No conclusions can be d awn ega ding noc u nal lying ac i i ies, as he e alua ion pe iod did no encompass nigh ime obse a ions. Lying beha io p edominan ly occu ed wi hin he clus e ed lying a ea, indica ing a p e e ence o speci ic lying zones. Addi ional spa ial dis ibu ion analysis could quan i y space equi emen s o all lying pigs, aking in o accoun hei endency o lie down simul aneously o a subs an ial po ion o he day. No ably, his s udy did no conside he lying pos u e o he pigs. Di e en ia ing be ween lying pos u es which a e known o a y h oughou he day [ 38 ] could p o ide addi ional insigh s in o clima e condi ions. 4.3. Beha io Sequences/P ocess Mining The ocus o his s udy was o in es iga e gene al pa e ns o beha io occu ence and succession a he han acking indi idual animals and hei comple e beha io sequence. The e a e a ious app oaches a ailable o indi idual acking, including hose based on ideo da a [ 46 ], as well as he u iliza ion o RFID (Radio F equency Iden i ica ion) echnol- ogy. RFID chips o e an al e na i e means o ack animal encoun e s and in e ac ions wi h objec s [ 47 – 49 ], which could p esen an in e es ing a enue as ID e i ica ion in combina ion wi h acking based on ideo da a. 4.3.1. De eca ing Pigs end o show speci ic beha io sequences like ea ing, d inking, u ina ing, and de eca ing as no mal beha io . The gene al de eca ion beha io o g ouped housed pigs in ol es sni ing be o e elimina ion in 50–70% o obse a ions and ollowed by mo ing away immedia ely a e elimina ion [ 42 ]. Typical elimina ion sequences a e explo ing- elimina ion-mo ing and mo ing-elimina ion-mo ing [ 43 ] o as desc ibed by Wechsle and Bachmann (1998) [ 50 ] in de ail: en e he dunging a ea, sni , pos u e, de eca e/u ina e. Sim- ila beha io al sequences can be obse ed in his s udy using p ocess mining o ideo da a. Clus e ed p ocess models iden i ied he sequence in es iga ing-de eca ing-in es iga ing oge he wi h in es iga ing-mo ing-in es iga ing as majo beha io sequences o he de eca ion p ocess which is consis en wi h he li e a u e. The analyses o de eca ion beha io and i s in eg a ion in o beha io sequences can se e as c ucial ea ly wa ning ools o de ec ing anomalies in managemen p ac ices [9,43] . De ia ions om he designa ed de eca ion a ea o al e a ions in de eca ion beha io wi hin beha io sequences due o ex e nal dis up ions o speci ic en i onmen al condi ions (e.g., clima ic ac o s) can be u ilized as aluable managemen indica o s. Moni o ing and ecognizing such changes in de eca ion pa e ns may aid in iden i ying po en ial issues in pig husband y and p o ide aluable insigh s o op imizing managemen s a egies o ensu e animal wel a e and o e all sys em e iciency. 4.3.2. Feeding The eeding beha io o a ening pigs is highly dependen on he managemen sys em. In eeding managemen s compa able o his s udy (wi h es ic ed access o eed and ixed eeding imes) pigs end o demons a e mo emen owa ds he eeding ough in close empo al p oximi y o hei egula eeding schedule. Addi ionally, pigs a e s imula ed by he g oup in mo emen and eeding [ 51 ]. Beha io sequences iden i ied in he clus e ed p ocess models in his s udy highligh he beha io sequence o mo ing- eeding, di ec link o eeding (Figu e 8a) o looped beha io o mo ing- eeding-mo ing (o in es iga ing- eeding-in es iga ing da a no shown he e) in be ween wo lying pe iods. A se e e amoun o eeding ac i i ies could be iden i ied be ween wo lying pe iods wi hou addi ional mo ing o in es iga ing beha io . These di e en indings gene ally comply wi h he assump ions om Signo e e al., (1975) [ 51 ]. This indica es ha he p ocess mining Ag icul u e 2023,13, 1639 17 o 20 app oach can iden i y ele an beha io al sequences. De ia ions a e no iceable in cases whe e eeding beha io di ec ly ollows lying beha io and hen ansi ions back in o lying beha io . I is highly likely ha his pa e n is in luenced by he du a ion o ime a speci ic beha io needs o be pe o med be o e i can be a ibu ed o a pa icula ca ego y. Fine- uning he h eshold se ings o his du a ion could be one possibili y o enhance he accu acy and obus ness o he models. The applica ion o p ocess mining on ideo da a enables he iden i ica ion o beha io sequences and holds po en ial o beha io al analysis unde a ious in luencing ac o s. Mo eo e , i can be conside ed as an ea ly wa ning ool, capable o de ec ing de ia ions om es ablished and ypical beha io sequences, acili a ing imely in e en ion and imp o ed managemen p ac ices. 5. Conclusions The s eps o he pa h om ac i i y ecogni ion in image da a o he new app oach o iden i ica ion o beha io al sequences using p ocess mining o desc ibing beha io al pa e ns in pigs exhibi simila app oaches o wha happens in he pig pen, albei wi h inc easing le els o de ail. All esul s om he s eps o he pa h demons a e plausible conclusions and can be measu ed wi h compa able app oaches. The p ocess models o desc ibing beha io al sequences in pig beha io ep esen a no el app oach ha success- ully iden i ies indi idual beha io al pa e ns and clus e s hem in o dis inc sequences. Fu u e in-dep h analyses could depic addi ional beha io al sequences wi h di e en s a and end beha io s o conside u he beha io al di e en ia ions. O e all, his me hod p esen s a p omising app oach o he au oma ed assessmen o pig beha io and beha - io al sequences based on ideo da a. The iden i ica ion o such beha io al sequences is c ucial o unde s anding he na u al low o ac i i ies and he unde lying pa e ns in pig beha io . This knowledge can in o m managemen decisions and shed ligh on po en ial de ia ions o anomalies ha equi e a en ion. The ecogni ion and spa ial assignmen o de eca ion beha io can be a aluable ool in educing wo kload and minimizing en i- onmen al impac by ensu ing adhe ence o designa ed de eca ion a eas, pa icula ly in eely en ila ed ba ns wi h s uc u ed unc ional a eas. Fu he in es iga ions a e needed o ans e his me hod o s uc u ed mul i-a ea pens wi h o wi hou eely en ila ed condi ions. Also addi ional e o s should ocus on e ining he p ocess mining app oach by explo ing di e en h eshold se ings and inco po a ing addi ional a iables o imp o e he models. Au ho Con ibu ions: Concep ualiza ion, A.M. and A.L.; me hodology, A.L. and J.B.; so wa e, A.L. and J.B.; alida ion, A.M., A.K. and E.H.; o mal analysis, A.M.; in es iga ion, A.M.; esou ces, A.M.; da a cu a ion, A.M, A.L. and J.B.; w i ing—o iginal d a p epa a ion, A.M. and A.L.; w i ing— e iew and edi ing, A.M., A.L. and A.K.; isualiza ion, A.L. and J.B.; supe ision, A.K. and E.H.; p ojec adminis a ion, A.M.; unding acquisi ion, A.M. and A.K. All au ho s ha e ead and ag eed o he published e sion o he manusc ip . Funding: This p ojec has ecei ed unding om he S a e o Schleswig-Hols ein unde he Da en- campus p ojec g an no. 220 21 016. Ins i u ional Re iew Boa d S a emen : E hical e iew and app o al we e wai ed o his s udy due o he na u e o he da a collec ion p ocess. The ideo da a eco ding was conduc ed in no mal a m condi ions, and no al e a ions o in e en ions we e made as pa o he s udy. As a esul , he e we e no po en ial isks o ha ms o human o animal subjec s, and he s udy s ic ly adhe ed o e hical guidelines and egula ions. Consequen ly, he animal wel a e o ice deemed ha an e hical e iew and app o al p ocess was unnecessa y o his pa icula s udy. Da a A ailabili y S a emen : The aw ideo ma e ials a e no publicly a ailable due o copy igh es ic ions. Da a including he esul s o objec de ec ion, objec acking, and ac i i y ecogni ion is a ailable unde he ollowing e e ence: [ 52 ]. The code can be downloaded unde he ollowing link: h ps://gi hub.com/a idle/ ideo-p ocess-mining-public (accessed on 18 Augus 2023). Ag icul u e 2023,13, 1639 18 o 20 Acknowledgmen s: We would like o hank he Schleswig-Hols ein Chambe o Ag icul u e, espe- cially he Fu e kamp expe imen al ba n, o he oppo uni y o eco d he ideos and o p o iding he accompanying pa ame e s. Con lic s o In e es : The au ho s decla e no con lic o in e es . Re e ences 1. Van Zels , S.J.; Mannha d , F.; de Leoni, M.; Koschmide , A. E en Abs ac ion in P ocess Mining: Li e a u e Re iew and Taxonomy. G anul. Compu . 2021,6, 719–736. [C ossRe ] 2. Chen, C.; Zhu, W.; No on, T. Beha iou Recogni ion o Pigs and Ca le: Jou ney om Compu e Vision o Deep Lea ning. Compu . Elec on. Ag ic. 2021,187, 106255. [C ossRe ] 3. Nasi ahmadi, A.; Hensel, O.; Edwa ds, S.A.; S u m, B. A New App oach o Ca ego izing Pig Lying Beha iou Based on a Delaunay T iangula ion Me hod. Anim. In . J. Anim. Biosci. 2017,11, 131–139. [C ossRe ] 4. Yang, Q.; Xiao, D.; Lin, S. Feeding Beha io Recogni ion o G oup-Housed Pigs wi h he Fas e R-CNN. Compu . Elec on. Ag ic. 2018,155, 453–460. [C ossRe ] 5. Fe nández-Ca ión, E.; Ma ínez-A ilés, M.; I o a, B.; Ma ínez-López, B.; Ramos, Á.M.; Sánchez-Vizcaíno, J.M. Mo ion-Based Video Moni o ing o Ea ly De ec ion o Li es ock Diseases: The Case o A ican Swine Fe e . PLoS ONE 2017 ,12, e0183793. [C ossRe ] 6. Be gamini, L.; Pini, S.; Simoni, A.; Vezzani, R.; Calde a a, S.; D’Ea h, R.; Fishe , R. Ex ac ing Accu a e Long-Te m Beha io Changes om a La ge Pig Da ase . In P oceedings o he 16 h In e na ional Join Con e ence on Compu e Vision, Imaging and Compu e G aphics Theo y and Applica ions, Vi ual, 8–10 Feb ua y 2021; SCITEPRESS—Science and Technology Publica ions: Se úbal, Po ugal, 2021; pp. 524–533. 7. Nannoni, E.; Aa nink, A.J.A.; Ve mee , H.M.; Reime , I.; Fels, M.; B acke, M.B.M. Soiling o Pig Pens: A Re iew o Elimina i e Beha iou . Anim. 2020,10, 2025. [C ossRe ] 8. Hacke , R.R.; Ogil ie, J.R.; Mo ison, W.D.; Kains, F. Fac o s A ec ing Exc e o y Beha io o Pigs. J. Anim. Sci. 1994 ,72, 1455–1460. [C ossRe ] 9. Höne, U.; K ause, E.T.; Bussemas, R.; T aulsen, I.; Sch ade , L. Usage o Ou doo Runs and De aeca ion Beha iou o Fa ening Pigs. Appl. Anim. Beha . Sci. 2023,258, 105821. [C ossRe ] 10. Aa nink, A.J.A.; an den Be g, A.J.; Keen, A.; Hoeksma, P.; Ve s egen, M.W.A. E ec o Sla ed Floo A ea on Ammonia Emission and on he Exc e o y and Lying Beha iou o G owing Pigs. J. Ag ic. Eng. Res. 1996,64, 299–310. [C ossRe ] 11. Ma hews, S.G.; Mille , A.L.; Clapp, J.; Plö z, T.; Ky iazakis, I. Ea ly De ec ion o Heal h and Wel a e Comp omises h ough Au oma ed De ec ion o Beha iou al Changes in Pigs. Ve . J. Lond. Engl. 1997 2016,217, 43–51. [C ossRe ] 12. Ma hews, S.G.; Mille , A.L.; Plö z, T.; Ky iazakis, I. Au oma ed T acking o Measu e Beha iou al Changes in Pigs o Heal h and Wel a e Moni o ing. Sci. Rep. 2017,7, 17582. [C ossRe ] 13. D 0 Ea h, R.B.; Jack, M.; Fu o, A.; Talbo , D.; Zhu, Q.; Ba clay, D.; Bax e , E.M. Au oma ic Ea ly Wa ning o Tail Bi ing in Pigs: 3D Came as Can De ec Lowe ed Tail Pos u e be o e an Ou b eak. PLoS ONE 2018,13, e0194524. [C ossRe ] 14. Oczak, M.; Viazzi, S.; Ismayilo a, G.; Sonoda, L.T.; Rouls on, N.; Fels, M.; Bah , C.; Ha ung, J.; Gua ino, M.; Be ckmans, D.; e al. Classi ica ion o Agg essi e Beha iou in Pigs by Ac i i y Index and Mul ilaye Feed Fo wa d Neu al Ne wo k. Biosys . Eng. 2014,119, 89–97. [C ossRe ] 15. Chen, C.; Zhu, W.; S eibel, J.; Sieg o d, J.; Wu z, K.; Han, J.; No on, T. Recogni ion o Agg essi e Episodes o Pigs Based on Con olu ional Neu al Ne wo k and Long Sho -Te m Memo y. Compu . Elec on. Ag ic. 2020,169, 105166. [C ossRe ] 16. Lepsien, A.; Koschmide , A.; K a sch, W. Analy ics Pipeline o P ocess Mining on Video Da a. In P oceedings o he BPM 2023 Fo um; Sp inge : Cham, Swi ze land, 2023; Volume 490. 17. Bundesminis e ium ü E näh ung und Landwi scha Ve o dnung Zum Schu z Landwi scha liche Nu z ie e Und Ande e Zu E zeugung Tie ische P oduk e Gehal ene Tie e Bei Ih e Hal ung (Tie schu z-Nu z ie hal ungs e o dnung—Tie SchNu z V): Tie SchNu z V. Bundesgese zbla 2021,2021, 142–145. 18. The Eu opean Pa liamen Council Di ec i e 2008/120/EC o 18 Decembe 2008 Laying down Minimum S anda ds o he P o ec ion o Pigs. Regul. EU 2017625 Eu . Pa liam. Counc. 15 Ma ch 2017. 2019. A ailable online: h ps://eu -lex.eu opa.eu/ legal-con en /EN/ALL/?u i=CELEX%3A32008L0120 (accessed on 18 Augus 2023). 19. Ve mee , H.; Di x-Kuijken, N.; B acke, M. Explo a ion Feeding and Highe Space Alloca ion Imp o e Wel a e o G owing- Finishing Pigs. Animals 2017,7, 36. [C ossRe ] 20. Spoolde , H.A.M.; Aa nink, A.A.J.; Ve mee , H.M.; an Riel, J.; Edwa ds, S.A. E ec o Inc easing Tempe a u e on Space Requi emen s o G oup Housed Finishing Pigs. Appl. Anim. Beha . Sci. 2012,138, 229–239. [C ossRe ] 21. Lepsien, A.; Bosselmann, J.; Mel sen, A.; Koschmide , A. P ocess Mining on Video Da a. In P oceedings o he ZEUS 2022, Vi ual, 24–25 Feb ua y 2022; Manne , J., Lübke, D., Haa mann, S., Kolb, S., He zbe g, N., Kopp, O., Eds.; CEUR-WS.o g: Bambe g, Ge many, 2022; Volume 3113, pp. 56–62. 22. Wang, C.-Y.; Bochko skiy, A.; Liao, H.-Y.M. YOLO 7: T ainable Bag-o -F eebies Se s New S a e-o - he-A o Real-Time Objec De ec o s. a Xi 2022, a Xi :2207.02696 1. Ag icul u e 2023,13, 1639 19 o 20 23. Deng, J.; Dong, W.; Soche , R.; Li, L.-J.; Li, K.; Fei-Fei, L. ImageNe : A La ge-Scale Hie a chical Image Da abase. In P oceedings o he 2009 IEEE Con e ence on Compu e Vision and Pa e n Recogni ion, Miami, FL, USA, 20–25 June 2009; pp. 248–255. 24. She min, T.; Teng, S.W.; Mu shed, M.; Lu, G.; Sohel, F.; Paul, M. Enhanced T ans e Lea ning wi h ImageNe T ained Classi ica ion Laye . In P oceedings o he PSVIT 2019, P oceedings 9, Sydney, NSW, Aus alia, 18–22 No embe 2019; Lee, C., Su, Z., Sugimo o, A., Eds.; Sp inge : Cham, Swi ze land, 2019; Volume 11854, pp. 142–155. 25. Zhang, Y.; Sun, P.; Jiang, Y.; Yu, D.; Weng, F.; Yuan, Z.; Luo, P.; Liu, W.; Wang, X. By eT ack: Mul i-Objec T acking by Associa ing E e y De ec ion Box. In P oceedings o he ECCV 2022, Tel A i , Is ael, 23–27 Oc obe 2022; A idan, S., B os ow, G., Cissé, M., Fa inella, G.M., Hassne , T., Eds.; Sp inge : Cham, Swi ze land, 2022; pp. 1–21. 26. Luo, W.; Xing, J.; Milan, A.; Zhang, X.; Liu, W.; Kim, T.-K. Mul iple Objec T acking: A Li e a u e Re iew. A i . In ell. 2021 , 293,103448. [C ossRe ] 27. Feich enho e , C.; Fan, H.; Malik, J.; He, K. SlowFas Ne wo ks o Video Recogni ion. In P oceedings o he ICCV 2019, Seoul, Republic o Ko ea, 27 Oc obe –2 No embe 2019; pp. 6201–6210. 28. Li, D.; Zhang, K.; Li, Z.; Chen, Y. A Spa io empo al Con olu ional Ne wo k o Mul i-Beha io Recogni ion o Pigs. Senso s 2020 , 20, 2381. [C ossRe ] 29. Kay, W.; Ca ei a, J.; Simonyan, K.; Zhang, B.; Hillie , C.; Vijayana asimhan, S.; Viola, F.; G een, T.; Back, T.; Na se , P.; e al. The Kine ics Human Ac ion Video Da ase . a Xi 2017, a Xi :1705.06950. 30. Zandka imi, F.; Rehse, J.-R.; Soudmand, P.; Hoehle, H. A Gene ic F amewo k o T ace Clus e ing in P ocess Mining. In P oceedings o he ICPM 2020, Padua, I aly, 5–8 Oc obe 2020; pp. 177–184. 31. Van de Aals , W.; Ad iansyah, A.; de Medei os, A.K.A.; A cie i, F.; Baie , T.; Blickle, T.; Bose, J.C.; an den B and, P.; B and jen, R.; Buijs, J.; e al. P ocess Mining Mani es o. In P oceedings o he BPM 2011 Wo kshops, Cle mon -Fe and, F ance, 29 Augus 2011; Daniel, F., Ba kaoui, K., Dus da , S., Eds.; Sp inge : Be lin/Heidelbe g, Ge many, 2012; Volume 99, pp. 169–194. 32. Shao, B.; Xin, H. A Real-Time Compu e Vision Assessmen and Con ol o The mal Com o o G oup-Housed Pigs. Compu . Elec on. Ag ic. 2008,62, 15–21. [C ossRe ] 33. Chung, Y.; Kim, H.; Lee, H.; Pa k, D.; Jeon, T.; Chang, H.-H. A Cos -E ec i e Pigs y Moni o ing Sys em Based on a Video Senso . KSII T ans. In e ne In . Sys . 2014,8, 1481–1498. [C ossRe ] 34. Cos a, A.; Ismayilo a, G.; Bo gono o, F.; Viazzi, S.; Be ckmans, D.; Gua ino, M. Image-P ocessing Technique o Measu e Pig Ac i i y in Response o Clima ic Va ia ion in a Pig Ba n. Anim. P od. Sci. 2014,54, 1075. [C ossRe ] 35. Bloemen, H.; Ae s, J.M.; Be ckmans, D.; Goedseels, V. Image Analysis o Measu e Ac i i y Index o Animals. Equine Ve . J. Suppl. 1997,29, 16–19. [C ossRe ] [PubMed] 36. Ni, J.-Q.; Liu, S.; Radcli e, J.S.; Vonde ohe, C. E alua ion and Cha ac e isa ion o Passi e In a ed De ec o s o Moni o Pig Ac i i ies in an En i onmen al Resea ch Building. Biosys . Eng. 2017,158, 86–94. [C ossRe ] 37. Cos a, A.; Ismayilo a, G.; Bo gono o, F.; Le oy, T.; Be ckmans, D.; Gua ino, M. The Use o Image Analysis as a New App oach o Assess Beha iou Classi ica ion in a Pig Ba n. Ac a Ve . B no 2013,82, 25–30. [C ossRe ] 38. Ekkel, E.D.; Spoolde , H.A.M.; Hulsegge, I.; Hops e , H. Lying Cha ac e is ics as De e minan s o Space Requi emen s in Pigs. Appl. Anim. Beha . Sci. 2003,80, 19–30. [C ossRe ] 39. Zo ic, M.; Johansson, S.-E.; Wallg en, P. Beha iou o Fa ening Pigs Fed wi h Liquid Feed and D y Feed. Po c. Heal h Manag. 2015,1, 14. [C ossRe ] 40. Pe sson, E.; Wülbe s-Minde mann, M.; Be g, C.; Alge s, B. Inc easing Daily Feeding Occasions in Res ic ed Feeding S a egies Does No Imp o e Pe o mance o Well Being o Fa ening Pigs. Ac a Ve . Scand. 2008,50, 24. [C ossRe ] 41. Bus, J.D.; Boumans, I.J.M.M.; Webb, L.E.; Bokke s, E.A.M. The Po en ial o Feeding Pa e ns o Assess Gene ic Wel a e in G owing-Finishing Pigs. Appl. Anim. Beha . Sci. 2021,241, 105383. [C ossRe ] 42. Ande sen, H.M.-L.; Kongs ed, A.G.; Jakobsen, M. Pig Elimina ion Beha io —A Re iew. Appl. Anim. Beha . Sci. 2020 ,222, 104888. [C ossRe ] 43. Guo, Y.; Lian, X.; Yan, P. Diu nal Rhy hms, Loca ions and Beha iou al Sequences Associa ed wi h Elimina i e Beha iou s in Fa ening Pigs. Appl. Anim. Beha . Sci. 2015,168, 18–23. [C ossRe ] 44. Tillmanns, M.; Scheepens, K.; S ol e, M.; He b and , S.; Kempe , N.; Fels, M. Implemen a ion o a Pig Toile in a Nu se y Pen wi h a S aw-Li e ed Lying A ea. Animals 2022,12, 113. [C ossRe ] [PubMed] 45. Ruckebusch, Y. The Rele ance o D owsiness in he Ci cadian Cycle o Fa m Animals. Anim. Beha . 1972 ,20, 637–643. [C ossRe ] 46. Van De Zande, L.E.; Guzh a, O.; Rodenbu g, T.B. Indi idual De ec ion and T acking o G oup Housed Pigs in Thei Home Pen Using Compu e Vision. F on . Anim. Sci. 2021,2, 669312. [C ossRe ] 47. Maselyne, J.; Ad iaens, I.; Huyb ech s, T.; De Ke elae e, B.; Mille , S.; Vangey e, J.; Van Nu el, A.; Saeys, W. Measu ing he D inking Beha iou o Indi idual Pigs Housed in G oup Using Radio F equency Iden i ica ion (RFID). Animal 2016 ,10, 1557–1566. [C ossRe ] [PubMed] 48. Gómez, Y.; S yga , A.H.; Boumans, I.J.M.M.; Bokke s, E.A.M.; Pede sen, L.J.; Niemi, J.K.; Pas ell, M.; Man eca, X.; Llonch, P. A Sys ema ic Re iew on Valida ed P ecision Li es ock Fa ming Technologies o Pig P oduc ion and I s Po en ial o Assess Animal Wel a e. F on . Ve . Sci. 2021,8, 660565. [C ossRe ] 49. Cappai, M.G.; Rubiu, N.G.; Pinna, W. Economic Assessmen o a Sma T aceabili y Sys em (RFID+DNA) o O igin and B and P o ec ion o he Po k P oduc Labelled “Suine o Di Sa degna”. Compu . Elec on. Ag ic. 2018,145, 248–252. [C ossRe ] Ag icul u e 2023,13, 1639 20 o 20 50. Wechsle , B.; Bachmann, I. A Sequen ial Analysis o Elimina i e Beha iou in Domes ic Pigs. Appl. Anim. Beha . Sci. 1998 , 56, 29–36. [C ossRe ] 51. Signo e , J.P.; Baldwin, B.A.; F ase , D.; Ha ez, E.S.E. The Beha iou o Swine. In Beha iou o Domes ic Animals; Bailliè e Tindall: London, UK, 1975; pp. 295–329. 52. Lepsien, A.; Koschmide , A.; K a sch, W. Video P ocess Mining E alua ion Da a. Zenodo 2023. [C ossRe ] Disclaime /Publishe ’s No e: The s a emen s, opinions and da a con ained in all publica ions a e solely hose o he indi idual au ho (s) and con ibu o (s) and no o MDPI and/o he edi o (s). MDPI and/o he edi o (s) disclaim esponsibili y o any inju y o people o p ope y esul ing om any ideas, me hods, ins uc ions o p oduc s e e ed o in he con en .