Dog behaviour classification with movement sensors placed on the harness and the collar

This is a sel -a chi ed e sion o an o iginal a icle. This e sion may di e om he o iginal in pagina ion and ypog aphic de ails. Au ho (s): Ti le: Yea : Ve sion: Copy igh : Righ s: Righ s u l: Please ci e he o iginal e sion: CC BY 4.0 h ps://c ea i ecommons.o g/licenses/by/4.0/ Dog beha iou classi ica ion wi h mo emen senso s placed on he ha ness and he colla © 2021 The Au ho s. Published by Else ie B.V. Published e sion Kumpulainen, Pekka; Ca dó, Anna Valldeo iola; Somppi, Sanni; Tö nq is , Heini; Vää äjä, Heli; Maja an a, Päi i; Giza dino a, Yulia; Hoog, An ink Ch is oph; Su akka, Veikko; Kujala, Miiamaa ia V.; Vainio, Ou i; Vehkaoja, An i Kumpulainen, P., Ca dó, A. V., Somppi, S., Tö nq is , H., Vää äjä, H., Maja an a, P., Giza dino a, Y., Hoog, A. C., Su akka, V., Kujala, M. V., Vainio, O., & Vehkaoja, A. (2021). Dog beha iou classi ica ion wi h mo emen senso s placed on he ha ness and he colla . Applied Animal Beha iou Science, 241, A icle 105393. h ps://doi.o g/10.1016/j.applanim.2021.105393 2021 Applied Animal Beha iou Science 241 (2021) 105393 A ailable online 1 July 2021 0168-1591/© 2021 The Au ho s. Published by Else ie B.V. This is an open access a icle unde he CC BY license (h p://c ea i ecommons.o g/licenses/by/4.0/). Dog beha iou classi ica ion wi h mo emen senso s placed on he ha ness and he colla Pekka Kumpulainen a , Anna Valldeo iola Ca d´ o b , Sanni Somppi b , Heini T¨ o nq is b , Heli V¨ a¨ a ¨ aj¨ a c , d , P¨ ai i Maja an a c , Yulia Giza dino a c , Ch is oph Hoog An ink e , Veikko Su akka c , Miiamaa ia V. Kujala b , , Ou i Vainio b , An i Vehkaoja a , * a Facul y o Medicine and Heal h Technology, Tampe e Uni e si y, P.O. Box 692, FI-33101, Tampe e, Finland b Depa men o Equine and Small Animal Medicine, Uni e si y o Helsinki, P.O. Box 57, FI-00014, Uni e si y o Helsinki, Finland c Resea ch G oup o Emo ions, Sociali y, and Compu ing, Facul y o In o ma ion Technology and Communica ion Sciences, Tampe e Uni e si y, P.O. Box 100, FI-33014, Tampe e, Finland d Lapland Uni e si y o Applied Sciences, Mas e School, Joki ¨ ayl¨ a 11 B, 96300, Ro aniemi, Finland e Biomedical Enginee ing (KIS*MED), TU Da ms ad , Magdalenens aße 2A, 64289 Da ms ad , Ge many Depa men o Psychology, Facul y o Educa ion and Psychology, Uni e si y o Jy ¨ askyl¨ a, PO Box 35, FI-40014, Uni e si y o Jy ¨ askyl¨ a, Jy ¨ askyl¨ a, Finland ARTICLE INFO Keywo ds: Dog Canine Beha iou classi ica ion Ac ig aphy Accele ome y Ac i i y moni o ing Wea able echnology ABSTRACT Dog owne s’ unde s anding o he daily beha iou o hei dogs may be enhanced by mo emen measu emen s ha can de ec epea able dog beha iou , such as le els o daily ac i i y and es as well as hei changes. The aim o his s udy was o e alua e he pe o mance o supe ised machine lea ning me hods u ilising accele - ome e and gy oscope da a p o ided by wea able mo emen senso s in classi ica ion o se en ypical dog ac- i i ies in a semi-con olled es si ua ion. Fo y- i e middle o la ge sized dogs pa icipa ed in he s udy. Two senso de ices we e a ached o each dog, one on he back o he dog in a ha ness and one on he neck colla . Al oge he 54 ea u es we e ex ac ed om he accele a ion and gy oscope signals di ided in wo-second seg- men s. The pe o mance o ou classi ie s we e compa ed using ea u es de i ed om bo h senso modali ies. and om he accele a ion da a only. The esul s we e p omising; he mo emen senso a he back yielded up o 91 % accu acy in classi ying he dog ac i i ies and he senso placed a he colla yielded 75 % accu acy a bes . Including he gy oscope ea u es imp o ed he classi ica ion accu acy by 0.7–2.6 %, depending on he classi ie and he senso loca ion. The mos dis inc ac i i y was sni ing, whe eas he s a ic pos u es (lying on ches , si ing and s anding) we e he mos challenging beha iou s o classi y, especially om he da a o he neck colla senso . The da a used in his a icle as well as he signal p ocessing sc ip s a e openly a ailable in Mendeley Da a, h ps://doi.o g/10.17632/ xhx934 bn.1. 1. In oduc ion Accele ome e s a e used in animal science in a ious con ex s a ying om he es ima ion o ene gy expendi u e (C ou e e al., 2006), assessmen o beha iou s in wildli e (Campbell e al., 2013; Halsey e al., 2011; Mo eau e al., 2009; Na han e al., 2012; Shepa d e al., 2008; Wilson e al., 2006) o applica ions in e e ina y medicine (Helm e al., 2016; Guillo e al., 2013; Li le e al., 2016). The numbe o consume - a ge ed ac i i y acke s a ailable o dogs has inc eased in ecen yea s and he ma ke is expec ed o g ow apidly in he o hcoming yea s. One o he key ac o s o he g ow h is dog owne s’ inc eased in e es and awa eness owa ds dog wellbeing. (G and View Resea ch, 2018) Combined wi h a sma phone applica ion, solu ions ypically isualize he da a as daily o al ac i i y, ype o ac- i i y (ligh o hea y), and collec i e beha iou s, such as he amoun o ime he dog has spen o mo ing o es ing du ing he day. The e is e idence ha exis ing ac i i y acke s a e easible o e alua ing simple canine beha iou s, o example, di e en ia ing be ween a seden a y ac i i y and wo in ensi ies o physical ac i i y (Yam e al., 2011). Depending on he sensi i i y o he measu emen uni , spon aneous ac i i ies o a dog, such as locomo ion, pos u al change and mo emen o body in each pos u e, can be di e en ia ed om accele ome e da a * Co esponding au ho . E-mail add ess: [email p o ec ed] (A. Vehkaoja). Con en s lis s a ailable a ScienceDi ec Applied Animal Beha iou Science jou nal homepage: www.else ie .com/loca e/applanim h ps://doi.o g/10.1016/j.applanim.2021.105393 Recei ed 17 Feb ua y 2021; Recei ed in e ised o m 7 June 2021; Accep ed 28 June 2021 Applied Animal Beha iou Science 241 (2021) 105393 2 (Yamada and Toku iki, 2000). Howe e , dog owne s and dogs could bene i om e en mo e accu a e and de ailed analysis o mo ion and body pos u es o he dog. De ailed de ec ion o dog’s e e yday ac i i ies would imp o e dog owne s’ unde s anding o pa icula dog beha iou s and eac ions, such as su e ing om sepa a ion anxie y while alone a home o in a kennel. Ex ac ing mo e de ailed beha iou s om accel- e ome e da a also has he po en ial o be used as an index o wellbeing and heal h s a us o he animal, o example, by de ec ing s ess and pain- ela ed beha iou s (Mo ison e al., 2014a; B own e al., 2010). Au oma ic beha iou dis inc ion would also bene i beha iou al esea ch, whe e beha iou s a e adi ionally measu ed by manual anno a ion o ideo eco dings, which is labou in ensi e and ime consuming and i would enable mo e de ailed beha iou al esea ch o ee- oaming wild animals (Ras e al., 2020). Some ac i i y logge s designed o humans ha e been commonly used o canine ac i i y moni o ing. These include “Ac iG aph GT3X / GTX3+” by Yam e al., 2011 and Mo ison e al., 2014b; as well as “Ac ical” by Hansen e al., 2007 and Olsen e al., 2016. These de ices a e p ima ily in ended o da a logging ( eco ding aw accele ome e da a) and hey do no classi y beha iou o mo ions o dogs au oma ically. In his s udy, we ha e de eloped and e alua ed he accu acy o beha iou classi ica ion om Ac iG aph da a using con en ional machine lea ning classi ie app oaches. In espec o accu acy o classi ica ion algo i hms de eloped o dogs, Ladha e al. (2013) achie ed a 68.6 % o global accu acy o di e en ia ing 16 canine beha iou s in na u alis ic en i- onmen s. den Uijl e al. (2017) showed ha walk, o , can e /gallop, ea , d ink, and headshake beha iou s could be classi ied wi h 95 % accu acy using a hie a chical “one s. he es ” classi ie . Addi ionally, se en canine ac i i ies such as si ing, and o ing we e dis inguished wi h an accu acy o mo e han 80 % by using 3-axis accele ome e and 3-axis gy oscope in (Ge encs´ e e al., 2013). Fe dinandy e al. (2020) ob ained up o 60 % o 80 % accu acy depending on he c oss- alida ion s a egy in classi ying eigh beha iou s and Ladha and Ho man (2018a) achie ed 86 % accu acy in de ec ing es ing o he dog. The placemen o he mo emen senso (e.g. neck o back) is an impo an ac o in he mo emen analysis. Rega ding e sa ili y and p ac icali y, he bes placemen o an ac i i y senso has been concluded o be en al a achmen o he neck colla , because his placemen makes i possible o de ec also beha iou s ha do no in ol e mo emen o he whole body, such as sc a ching and ea ing (Hansen e al., 2007). On he o he hand, accele ome e s a ached a he back may be able o di e en ia e beha iou s ha de ices a ached o he colla canno de ec such as ele a ed walking eloci y (P es on e al., 2012). The aim o his s udy was o e alua e ou commonly used classi i- ca ion algo i hms o dis inguishing se en dog beha iou s using ine ial senso da a eco ded in semi-con olled es si ua ions. The beha iou s we e galloping, lying on ches , si ing, sni ing, s anding, o ing, and walking. Fu he mo e, as some beha iou al phenomena may be be e de ec able om he dog’s neck o back loca ions, we sys ema ically examined how he de ice placemen , ei he on he ha ness a he back o on he colla a ound he neck, a ec ed he classi ica ion accu acy. Ini ial esul s o he s udy wi h smalle numbe o dogs and wi h only one senso loca ion (neck) ha e ea lie been published in ACI (Animal Compu e In e ac ion) con e ence (Kumpulainen e al., 2018). 2. Ma e ials and me hods 2.1. Tes se up The expe imen s we e conduc ed a he Uni e si y o Helsinki, Fac- ul y o Ve e ina y Medicine. The s udy p o ocol was e iewed and accep ed by he E hical Commi ee o he Use o Animals in Expe i- men s a he Uni e si y o Helsinki (minu es 5/2017). All dog owne s signed an in o med consen be o e pa icipa ing in he s udy. The a - endan s we e ee o cancel hei pa icipa ion a any ime wi hou gi ing a eason. A o al o 45 heal hy, middle o la ge -sized pe dogs om 27 b eeds pa icipa ed in he s udy. The a e age age o he dogs was 4.9 yea s ( ange 1–9 yea s) and he a e age weigh was 24.5 kg ( ange 13–41 kg). Table 1 shows de ailed b eed, weigh and age s a is ics o he pa ici- pa ing dogs. The da ase is desc ibed in mo e de ail in a sepa a e da ase a icle (Vehkaoja e al., 2021a) and is eely a ailable in Mendeley Da a (Vehkaoja e al., 2021b). 2.2. Tes p o ocol The es s we e conduc ed in a dog spo ing hall in a es ing a ena o 10m ×18m co e ed wi h a i icial u . The es sequence consis ed o se en asks whe e he owne was ins uc ed o guide he dog acco d- ingly. Th ee o he asks we e s a ic asks (i.e. si ing, s anding, lying down) and ou we e dynamic asks (i.e. o ing, walking, playing, and ea - sea ching), each ask las ed o h ee minu es. The whole p ocedu e was epea ed a e a sho b eak while changing he o de o he asks. Dogs pe o med asks sequen ially, al e na ing be ween s a ic and dynamic asks. T ea sea ch was always pe o med as he inal ask o he sequence and i consis ed o sea ching small pieces o d y dog ood sp ead on he g ound (a ea o 4m ×4m) by sni ing. Dogs wo e wo Ac iG aph GT9X Link (Ac iG aph LLC, Flo ida, USA) ac i i y senso s including 3-axis accele ome e and 3-axis gy oscope senso s (sampling a e 100 Hz). One senso was placed inside a igh pocke made o neop ene on he back bel o he dog’s ha ness, e e ed o as he back senso in his pape . The o he senso was a ached igh ly wi h an adhesi e ape on he en al side o he neck colla and is e e ed o as he colla senso in his pape . Dogs we e on he leash (1.5 m) and we e led by hei owne s o he expe imen e . The leash was connec ed o a sepa a e colla ha was placed close o he dogs’ body han he colla o which he senso was a ached. The owne s we e allowed o gi e ood ewa ds and command hei dogs h ough he en i e es . 2.3. Beha iou anno a ion The ac ual beha iou o he dogs du ing he assigned asks we e anno a ed using ideo eco dings. The es p ocedu e was eco ded wi h Table 1 Cha ac e is ics o he 45 dogs ha pa icipa ed in he s udy. B eed Numbe Weigh (kg) Age (yea s) Aus alian Kelpie 1 18 3 Beauce Shephe d 3 30.33 (28–35) 3 (3 – 3) Belgian Shephe d 1 29 6 Belgian Shephe d G oenendael 1 20 5 Belgian Shephe d Malinois 1 25 3 Bo de Collie 4 16.5 (15–20) 3.75 (3–5) Bou ie des Fland es 1 30 7 Bou ie des A dennes 2 22.5 (22–23) 4.5 (4–5) Bull Te ie (Minia u e) 1 17 2 C ossb eed 4 16.25 (13–20) 4.5 (3–7) Du ch Shephe d 2 24 (23–25) 3 (3−3) English Sp inge Spaniel 1 25 4 Finnish Lapphund 1 26 5 Fla -Coa ed Re ie e 1 28 4 Ge man Shephe d 3 32.33 (30–35) 3 (3−3) Golden Re ie e 3 30 (23–39) 4.67 (4–5) Ho awa 2 34.5 (28–41) 5 (5−5) Lab ado Re ie e 3 30 (23–37) 3 (3−3) Lago o Romagnolo 1 14 7 Lapponian He de 2 21 (20–22) 3 (3−3) Mudi 1 16 7 No a Sco ia Duck Tolling Re ie e 1 20 5 Smoo h Collie 1 18 3 Spanish Wa e Dog 2 22.5 (20–25) 7 (7−7) S anda d Poodle 1 31 7 Hunga ian Vizsla 1 25 2 P. Kumpulainen e al. Applied Animal Beha iou Science 241 (2021) 105393 3 Panasonic HDC-SD600 and Sony HDR-CX450 ideo came as posi ioned on he opposi e la e al walls and acing owa ds he es ing a ena. The pos hoc anno a ion o he ideo eco dings was done using he Obse e XT 10.5 so wa e (Noldus, The Ne he lands). Only segmen s longe han one second we e included in he anno a ion. Dynamic beha iou s (i.e. Walking, T o ing, Galloping, Sni ing; Table 2) we e only encoded i unambiguous, i.e. i he e was only one ob ious, con inuous dynamic beha iou wi hou he dog leaning owa ds he handle o pulling he leash, hus a ec ing he gai pa e n o he body posi ion. Galloping was anno a ed only du ing he play ask and sni ing du ing he ea sea ch ask (see Table 2 o he e hog am). S a ic beha iou s consis ed o s ill pos u es (i.e. Lying on ches , Si ing, S anding) and anno a ed when limbs did no mo e and he e was no physical con ac be ween he handle and he dog, excep i a ea was gi en. 2.4. Fea u e ex ac ion and labelling The aw ime se ies da a p oduced by he mo emen senso s we e sa ed wi h Ac iLi e so wa e (Ac iG aph LLC, Flo ida, USA) and ana- lysed o line wi h MATLAB R2018b (The Ma hWo ks, Inc., Na ick, MA USA). The ime se ies signals we e segmen ed in o wo-second ime win- dows wi h 50 % o e lap. A o al o 27 ea u es pe senso ype we e calcula ed o each segmen and used o classi ica ion o he beha - iou s. The same ea u es we e used o bo h accele ome e (A1 – A27) and gy oscope (G1 – G27) da a. The desc ip ions o he esul ing 54 ea u es a e gi en in Table 3. The in e pola ed in e se empi ical cumula i e dis ibu ion unc ion (ecd ) has been p esen ed by Cox and Oakes (1984) and i has ea lie been used o compu ing mo emen ela ed ea u es by Hamme la e al. (2013). The ecd ea u es a e based on he cumula i e dis ibu ion unc ion P c (x) =P(X ≤x) in a ollowing way. Se en alues o p i , e enly dis ibu ed be ween 0 and 1, in each x, y and z axis we e selec ed. Fo each p i he alue x i o which P(X ≤x i ) =p i was es ima ed by shape-p ese ing piecewise cubic in e pola ion. Fig. 1 le panel shows an example o cumula i e dis ibu ion unc ion o a no mal dis ibu ion. The igh panel shows wo examples de i ed om wo-second windows o accele ome e signal in x di ec ion du ing walking and o ing. The ue beha iou classes (see Table 2) o he da a segmen s we e assigned acco ding o he ideo anno a ions and synch onized o ma ch he imes amps in he ideo anno a ions. A beha iou class was assigned as a label o a segmen i a single anno a ed beha iou was occu ing a minimum o 75 % o he segmen . This was done in o de o inc ease he amoun o included da a, especially o he beha iou s ha ypically occu in sho du a ions, namely sni ing and galloping. The da a o bo h es sequences we e included o 17 dogs and only one es sequence o he emaining 28 dogs. The eason o including only one o he es sequences o he 28 was he challenges aced in eliable synch oniza- ion o he da a. The 62 es s om he 45 dogs p o ided 54,594 ins ances o labelled da a. Table 4 shows how he segmen s we e dis ibu ed be- ween he beha iou s. 2.5. Fea u e selec ion and classi ica ion All 54 ea u es we e Z-sco e no malised o ze o mean and uni a iance. Due o he high numbe o ea u es, he ea u e selec ion was pe o med in wo pa s. Fi s , weigh s o impo ance we e calcula ed o each ea u e by he Relie F algo i hm (Robnik-Sikonja and Kononenko, 2003) based on he k-nea es neighbou app oach. In o de o educe compu a ional cos s, only he mos impo an ea u es we e used in he subsequen o wa d ea u e selec ion. He e, ea u es a e added one a a ime in an o de in which hey bes imp o e he classi ica ion accu acy. This is con inued un il addi ional ea u es p o ide no imp o emen . The o wa d selec ion was pe o med o each o he ou classi ie s. Bo h, he o wa d selec ion and he inal classi ica ion esul s we e compu ed using Lea e-One-Dog-Ou c oss alida ion. Thus, all he da a o each indi idual dog we e le ou a a ime as a es se and he classi ie s we e iden i ied using he es o he da a. Lea e one subjec ou c oss ali- da ion has been p oposed o e he andom spli ing o he da a in which case he da a om he same indi idual easily ends up in bo h aining and alida ion da a se s p oducing o e ly goods classi ica ion esul s. This was ecen ly shown by Fe dinandy e al. (2020) in he con ex o animal beha iou classi ica ion. The whole ea u e selec ion p ocedu e was epea ed using only he accele ome e ea u es o e i y he bene i o he in o ma ion p o ided by he gy oscopes. The ou classi ie s in his s udy we e linea and quad a ic disc im- inan analysis classi ie s (LDA and QDA, espec i ely), a suppo ec o machine (SVM) classi ie wi h gaussian ke nel, and a classi ica ion ee (Duda e al., 2000). The egula isa ion o he classi ica ion ee was con olled by he numbe o cu s allowed in he ee. The numbe was op imised wi h he c oss alida ion using he mos signi ican ea u es gi en by he elie weigh s and ha alue was used in he o wa d ea u e selec ion he ea e . These ou classi ie s we e chosen based on hei popula i y in basic machine lea ning s udies and hei simple s uc u e and low compu a ional cos ha would enable hei in eg a- ion also in o a powe cons ained embedded measu emen pla o m in he u u e. The p edic ed class was ob ained as he highes p obabili y class p oposed by he classi ie e e ed o as o e all accu acy in (Fe - dinandy e al., 2020). 2.6. S a is ical analysis The classi ica ion accu acies epo ed in he esul s we e calcula ed Table 2 E hog am o he beha iou s included in he s a is ical analyses. Beha iou Desc ip ion Galloping 3- o 4-bea gai whe e he dog li s and pu s down bo h on and ea ex emi ies in a coo dina ed manne , in 1−2-3-bea gai (can e ) o in 1−2-3−4 bea gai (gallop). All ou ex emi ies a e simul aneously in he ai a some poin in e e y s ide. Galloping occu ed only du ing Playing ask. Lying on ches The dog’s o so is ouching he g ound and hips a e in he same le el as shoulde s. The dog can change balance poin wi hou using limbs. Si ing The dog has ou ex emi ies and ump on he g ound. The dog can change balance poin om cen al o hip o ice e sa. Sni ing The dog has i s head below i s back line and mo es i s muzzle close o he g ound. The dog walks, s ands o pe o ms ano he slow mo emen , bu i s ches and bo om do no ouch he g ound. Taking ood om he g ound and ea ing i can be included (ea ing was no coded sepa a ely). S anding The dog has he ou ex emi ies on he g ound, wi hou he dog’s o so ouching he g ound. T o ing 2-bea gai whe e he dog li s and pu s down ex emi ies in diagonal pai s a a speed as e han walking. Walking 4-bea gai whe e he dog mo es ex emi ies a slow speed, legs a e mo ed one by one in he o de : le hind leg, le on leg, igh hind leg, and igh on leg. The dog mo es s aigh o wa d o a maximum in 45 deg ees angle. Table 3 Desc ip ion o he ea u es calcula ed o 2-second segmen s o ime se ies mo emen senso da a. “A” e e s o accele ome e and “G” o gy oscope. Fea u e code Fea u e desc ip ion A1, G1 To al ac i i y: sum o s anda d de ia ion in all h ee axis A2, G2 Posi ion o se : Euclidean dis ance om he obus mean ob ained while he dogs we e s anding s ill A3, G3 The numbe o mean c ossings, he sum o x, y and z axis A4 – A6, G4 – G6 The mean alue o each axis; x, y, z A7 – A27, G7 – G27 In e pola ed in e se empi ical cumula i e dis ibu ion unc ion (ecd ): se en alues o each axis, a o al o 21 ea u es o each senso ype P. Kumpulainen e al. Applied Animal Beha iou Science 241 (2021) 105393 4 as he a e ages o he pe cen ages o co ec ly p edic ed beha iou class in all olds o he c oss alida ion. Di e ences be ween he classi ie s we e es ed a p =0.05 le el using - es , which assumes no mal dis- ibu ion. The no mali y o he classi ica ion a e dis ibu ion o each classi ie was es ed by Kolmogo o -Smi no es . The s a is ical ana- lyses we e conduc ed by MATLAB S a is ics and Machine Lea ning Toolbox R2018b. 3. Resul s The elie ea u e weigh s we e calcula ed wi h six alues o k: {3, 5, 9, 13, 17, 21}. Fo he ea u e se wi h bo h accele ome e and gy oscope da a, he ea u es included in he op 20 weigh s by any o he k alues (23 ea u es o he back and 22 o he colla senso ) we e used in he o wa d selec ion phase. When e alua ing he accele ome e da a alone, he ea u es we e selec ed in he same way bu choosing he ea u es included in he op 15 weigh s by any o he k alues (17 ea u es o he back and 15 o he colla senso ). The op imal numbe s o cu s acqui ed o he classi ica ion ees we e 124 o he back and 121 he colla senso o all ea u es, and 168 o he back and 54 o he colla senso o he accele ome e ea u es only. The ea u es selec ed by he o wa d selec ion and he inal c oss- alida ed classi ica ion accu acies a e p esen ed in Table 5. In all cases, mo e accele ome e ea u es and ewe gy oscope ea- u es we e selec ed. Conside ing also he gy oscope ea u es in he classi ica ion p o ided be e accu acy wi h all classi ie s and bo h senso loca ions. The e o e, all he de ailed esul s p esen ed below a e p esen ed o he cases wi h he selec ed accele ome e and gy oscope ea u es. Con usion ma ices o he classi ica ion esul s a e p esen ed in Fig. 2. The mos challenging beha iou s o classi y wi h he da a o he colla senso we e he s a ic pos u es: lying on ches , si ing, and s anding. Lying on ches was mos o en mixed wi h he o he s a ic pos u es. The back senso p o ided simila esul s bu he e he di e - ence in he accu acies be ween he classes was no so clea . The ac i i ies ha in ol ed mo emen we e gene ally classi ied e y accu a ely, mos ly highe han 90 %, sni ing being he mos dis inc beha iou in bo h senso loca ions and almos all classi ie s. Walking was classi ied wi h sligh ly wo se accu acy wi h he neck senso and was mixed wi h s a ic beha iou s. The classi ica ion esul s we e also calcula ed sepa a ely o each dog o s udy he di e ences in he accu acy be ween indi idual dogs. The esul s a e shown as boxplo s in Fig. 3. The box con ains he in e qua ile ange be ween he 25 h and 75 h pe cen iles. The no ch a ound he median co e s 95 % con idence limi s. The whiske s ex end o he ex eme da a poin up o 1.5 imes he in e qua ile ange om he box. Indi idual poin s ou side he maximum leng h o he whiske s a e ma ked wi h ed c osses. Ma king he con idence in e al o he median alue allows isual inspec ion o he s a is ically signi ican di e ences be ween he esul s ob ained wi h di e en classi ie s. I he no ches o he esul s o wo classi ie s do no o e lap, he medians a e s a is ically di e en . As seen in Fig. 3, he esul s a y conside ably be ween he dogs. The highes classi ica ion accu acies o some indi idual dogs a e abo e 99 Fig. 1. Illus a ion o ecd ea u e calcula ion. Le panel: one poin om no mal dis ibu ion a P(X ≤x) =p i . Righ panel: example o ac ual da a e alua ed a se en p i alues be ween 0 and 1. Table 4 The numbe o segmen s assigned o each beha iou . Beha iou Coun % o o al Galloping 776 1.4 Lying on ches 11,062 20.3 Si ing 10,824 19.8 Sni ing 9331 17.1 S anding 9552 17.5 T o ing 6646 12.2 Walking 6403 11.7 Table 5 Classi ica ion accu acies and he selec ed ea u es o each classi ie and bo h ea u e se scena ios sepa a ely o senso s loca ed on he back and he neck. Back senso Colla senso Classi ie Measu emen Accu acy % Fea u es Accu acy % Fea u es LDA A +G 89.0 9: A{3, 15, 18, 19, 25, 26} 71.9 10: A{5, 6, 15, 18, 19, 22, 26} G{6, 22, 26} G{6, 22, 27} A only 87.8 8: A{3, 15, 18, 19, 23, 24, 25, 26} 71.1 9: A{3, 5, 6, 15, 16, 18, 19, 23, 26} QDA A +G 91.1 9: A{2, 8, 17, 19, 26} 70.6 11: A{5, 6, 15, 17, 19, 22, 23, 26} G{5, 6, 22, 25} G{6, 22, 27} A only 90.2 5: A{8, 11, 17, 19, 26} 69.5 6: A{5, 6, 15, 19, 23, 26} SVM A +G 91.4 8: A{6, 8, 15, 18, 19, 24} 75.6 10: A{3, 5, 6, 15, 18, 19, 22, 26} G{22, 26} G{22, 27} A only 90.5 6:A{6, 8, 15, 18, 24, 26} 74.9 7: A{3, 5, 6, 15, 17, 19, 26} T ee A +G 91.0 10: A{5, 8, 9, 15, 18, 19} 72.4 7: A{3, 5, 6, 17, 19, 26} G{27} G{5, 6, 22, 25} A only 88.7 6: A{6, 8, 15, 17, 18, 25} 69.8 9: A{3, 5, 6, 15, 17, 19, 23, 24, 26} P. Kumpulainen e al. Applied Animal Beha iou Science 241 (2021) 105393 5 % o all classi ie s wi h he senso a ached on he ha ness. The colla senso eaches 90 % accu acy o some dogs. Howe e , he lowes ac- cu acies a e be ween 47 % and 66 %. Fo he back senso , di e ences in he dis ibu ions o he esul s o indi idual dogs ob ained wi h di e en classi ie s we e no s a is ically signi ican . Fo he colla senso , SVM ga e signi ican ly be e esul s han LDA and QDA. LDA and QDA we e no signi ican ly di e en om each o he . The esul s ob ained wi h he classi ica ion ee we e no signi ican ly di e en om any o he o he classi ie s. Table 6 p o ides all pai wise p- alues o he dog-wise accu acy dis ibu ions o di e en classi ie s. 4. Discussion Consume - a ge ed dog ac i i y me e s a e widely a ailable on he ma ke , bu he in o ma ion hey gi e o he dog owne s is a he limi ed. The aim o his s udy was o e alua e he pe o mance o ac i i y classi ica ion wi h wo mo emen senso s loca ed in he colla and he ha ness. The senso s p o ided bo h accele ome e and gy o- scope da a o classi ying se en ac i i ies o dogs in a semi-con olled es si ua ion. The esul s we e p omising, yielding up o 91 % Fig. 2. Con usion ma ices o he ou classi ie s. The ue classes a e in he ows and he p edic ed classes in he columns. Fig. 3. Boxplo s o he accu acies o he ou classi ie s (LDA, QDA, SVM, T ee). Each box con ains he classi ica ion accu acies o he 45 dogs. Table 6 p- alues o he pai wise - es s be ween he classi ie s. The alues o he back senso a e in he lowe le and he colla senso in he uppe igh pa . Bold- aced alues show s a is ical signi ican di e ence (p <0.05) be ween he classi ie s in dog-wise classi ica ion accu acy dis ibu ions. Colla senso LDA QDA SVM T ee LDA 0.428 0.049 0.786 QDA 0.289 0.008 0.318 SVM 0.208 0.752 0.112 T ee 0.241 0.886 0.856 Back senso P. Kumpulainen e al. Applied Animal Beha iou Science 241 (2021) 105393 6 classi ica ion accu acy using he da a o bo h senso ypes om he senso a he back. Including he gy oscope da a in addi ion o he accele ome e da a p o ided 0.7 %–2.6 % be e accu acy wi h all ou classi ie s and bo h senso loca ions. Sni ing was he mos dis inc beha iou , esul ing in 99.2 % accu- acy wi h he colla senso and 98.0 % wi h he back senso . This is in con as o an ea lie inding by Ladha e al. (2013) who ound ha walking and unning (o he beha iou s sha ed wi h his wo k) had be e classi ica ion pe o mance han sni ing. Howe e , all classes included in a classi ica ion ask a ec he pe o mance o each indi- idual class, which may explain he di e ence. The mos challenging ask o he classi ie s was di e en ia ion be ween he s a ic pos u es wi h he colla senso , namely lying down, si ing, and s anding. F om hose ime segmen s whe e he dog was lying down, only 28%–45% we e classi ied co ec ly, and he es we e classi ied mainly as ei he si ing o s anding. Howe e , conside ing he mino o ien a ion change in he dogs’ neck and back du ing hese asks, he mixing o hese pos- u es is a he logical. den Uijl e al. (2017) epo simila esul s, while hey had he lowes speci ici y o sleep beha iou . Howe e , hey did no classi y s anding, si ing, and lying down sepa a ely, bu a combined class as s a ic/inac i e, which makes he classi ica ion ask signi ican ly easie . In gene al, he esul s showed ha ac i i y moni o s on he back yielded be e esul s o classi ica ion han a aching an ac i i y moni o on he neck colla . The esul s o he back senso a e in line wi h hose by Ge encs´ e e al. (Ge encs´ e e al., 2013) who also used a senso a he back. The placemen o he senso has been shown o a ec he amoun o measu ed ac i i y in p e ious s udies (Hansen e al., 2007; P es on e al., 2012). In addi ion, igh ness o he a achmen may a ec he accu acy. Fo example, P es on e al. (2012) ound ha accele ome e s a ached igh ly o he back de ec ed beha iou s mo e accu a ely han de ices a ached loosely o he back. In ou s udy, he a achmen echnique o he senso s was dependen on he placemen : he colla senso was a ached o he colla wi h adhesi e ape, bu he back senso was inse ed in a neop ene pocke . This migh be one eason o he di e - ence be ween he esul s o he colla and he back senso s. Al hough he pocke was igh , he mo emen s o he senso could be a ec ed by i , as has been ea lie concluded by Ma in e al. (Ma in e al., 2016). In ou s udy, a likely eason o he di e ence in he classi ica ion accu acy be ween he colla and he back senso s is ha he dis inc s a ic pos- u es esul in mo e signi ican changes in he o ien a ion o a senso a ached o he back han o he colla . Ano he likely eason is ha , because he o ien a ion o he neck senso may change sligh ly due o u ning o he colla a ound he neck, he o ien a ions eco ded in he s a ic asks may o e lap. Following his, he po en ial o a ion o he colla needs o be conside ed and compensa ed as has also been concluded in (Ladha e al., 2018b). Walking was classi ied wi h high accu acy by he back senso bu mixed wi h s a ic pos u es wi h he neck senso . I should be no ed ha walking in con olled es si ua ion was a he di e en om eal li e, whe e a dog a ely pu ely walks slowly in leash, bu a he mixes walk and pace gai s. As he classi ica ion accu acies a y conside ably be ween indi idual dogs and he manually anno a ed ideo da a om e e y dog does no necessa ily con ain he same amoun o all beha iou s. Thus, he ques- ion a ises whe he he p opo ions o he beha iou s ha e an e ec on he indi idual classi ica ion esul s. Howe e , no signi ican co ela ion was ound be ween he class p opo ions and he classi ica ion a e, excep o a nega i e co ela ion o walking beha iou wi h QDA and SVM classi ie s on he back senso . Howe e , walking was one o he bes classi ied beha iou s in his s udy, so his co ela ion mus ha e happened by a pu e chance. We es ed only medium o la ge sized dog b eeds o ge a homologous pa icipan g oup wi h smalle a iabili y. Fo example, Ladha e al. (2013) ound be e global accu acy o di e en ia ing beha iou s in small and medium sized dogs han in la ge ones. In he u u e, he accu acy o he cu en beha iou classi ica ions should be es ed in a la ge dog popula ion ha includes also smalle dogs as well as dogs wi h de ian body s uc u e (e.g. sho legs). As ou es se up was also ela i ely con olled, u u e s udies should also be done wi h na u al beha iou s o dogs mo ing eely in hei amilia en i onmen . Al hough he senso was a ached o a di e en colla han he leash, he handle ’s beha iou may ha e a ec ed he senso , especially in hose si ua ions when he dog mo ed slowe han he handle did. I would be ideal i he leash could be a ached o he ha ness and he senso o he colla (o ice e sa), bu as we aimed o es bo h senso placemen s simul aneously, his was no an op ion. In u u e s udies, o eliable beha iou de ec ion, he de ice should always be posi ioned a exac ly he same o ien a ion, o he o ien a ion should be ecalib a ed a e each ime i is a ached (as was done in he p esen s udy) as well as each ime i may ha e been shi ed. SVM classi ie p o ides he bes esul s o bo h senso s, bu he di e ence in pe o mance is s a is ically signi ican only o he colla senso . In all cases, he classi ica ion a es ha e high de ia ion be ween indi idual dogs. Thus, o p ac ical pu poses, uning he classi ie o each indi idual would be bene icial om he accu acy poin o iew bu may no be necessa ily easible in p ac ice. Including he gy oscope da a in he classi ica ion lowe s he misclassi ica ion a es. Howe e , o p ac ical embedded p oduc s, i is a comp omise whe he he imp o e- men is wo h he added complexi y and dec eased ba e y li e ime. 5. Conclusion Ou cu en esul s sugges ha beha iou classi ica ion was mo e success ul om he mo emen senso a ached o he ha ness a he back o he dog a he han on he neck colla . In pa icula , s a ic beha iou s o si ing, s anding, and lying down we e ha d o di e en ia e wi h he senso a ached o he colla . Posi ioning may comp ise a challenge o he usabili y o ac i i y moni o s o di e en ia ing beha iou s in eal li e. A aching he senso o he colla is con enien o he dog and he owne , bu i i comp omises di e en ia ion o es ing om o he seden a y beha iou s as concluded in (den Uijl e al., 2017), i can lead o misleading conclusions in cases whe e es beha iou is used, o example, as an indica o o a dog’s pain o s ess le el. Ou cu en esul s a e p omising in e ms o de elopmen o p ac- ical me hods o au oma ically gaining in o ma ion on dog beha iou . This ype o mo e accu a e in o ma ion can be use ul in suppo ing he owne in gaining o e all unde s anding o a dog’s daily li e, assessmen o heal h o sickness, and unc ioning o medica ion, in pa icula o dogs su e ing om ch onic illness. In he u u e, he echnique could be de eloped u he o iden i y beha iou p oblems, hei causes, hei ea men as well as he e ec i eness o he ea men and assess he issues and changes in o e all wel a e based on he da a. Fu he mo e, ou p esen esul s pa e he way o de eloping solu ions o associa e he ac i i y o he a ec i e s a e o he dog, o suppo a mo e comp e- hensi e assessmen o dog wel a e. Au ho con ibu ions Concep ualiza ion: PK, SS, HT, PM, VS, MVK, OV, AV. Me hodology: PK, AVC, SS, HT, MVK, AV. So wa e: PK, HV, YG, AV. Fo mal Analysis: PK, AV. In es iga ion: AVC, SS, HT. Da a Cu a ion: PK, AV. W i ing – O iginal D a : PK, AV. W i ing – Re iew & Edi ing: PK, AVC, SS, HT, HV, PM, YG, CHA, VS, MVK, OV, AV. Visualiza ion: PK, AV. Funding Acquisi ion: PM, VS, OV, AV. Da a a ailabili y The anno a ed da ase s gene a ed and used in his s udy as well as he amewo k desc ip ion and sc ip s used in da a analysis a e a ailable in a eposi o y hos ed by Mendeley Da a. The DOI o he da ase is: h ps://doi.o g/10.17632/ xhx934 bn.1. The da a is desc ibed in mo e P. Kumpulainen e al. Applied Animal Beha iou Science 241 (2021) 105393 7 de ail in he ela ed Da a in B ie a icle: Mo emen Senso Da ase o Dog Beha io Classi ica ion. Decla a ion o Compe ing In e es The au ho s decla e no compe ing in e es s. Acknowledgemen s The au ho s would like o hank all he pa icipan s o he s udy. This esea ch was unded by Business Finland, a Finnish unding agengy o inno a ion, g an numbe s 1665/31/2016, 1894/31/2016, 7244/31/ 2016 in he con ex o “Buddy and he Smi hs 2.0” p ojec . Re e ences B own, D.C., Bos on, R.C., Fa a , J.T., 2010. Use o an ac i i y moni o o de ec esponse o ea men in dogs wi h os eoa h i is. J. Am. Ve . Med. Assoc. 237 (1), 66–70. h ps://doi.o g/10.2460/ja ma.237.1.66. Campbell, H.A., Gao, L., Bidde , O.R., Hun e , J., F anklin, C.E., 2013. C ea ing a beha iou al classi ica ion module o accele a ion da a: using a cap i e su oga e o di icul o obse e species. J. Exp. Biol. 216 (24), 4501–4506. h ps://doi.o g/ 10.1242/jeb.089805. Cox, D.R., Oakes, D., 1984. Analysis o Su i al Da a. Chapman & Hall, London. h ps:// doi.o g/10.1201/9781315137438. C ou e , S., Chu illa, J., Basse , D., 2006. Es ima ing ene gy expendi u e using accele ome e s. Eu . J. Appl. Physiol. 98, 601–612. h ps://doi.o g/10.1007/ s00421-006-0307-5. den Uijl, I., G´ omez ´ Al a ez, C.B., Ba am, D., D o , Y., Holland, R., Cook, A., 2017. Ex e nal alida ion o a colla -moun ed iaxial accele ome e o second-by-second moni o ing o eigh beha iou al s a es in dogs. PLoS One 12 (11) h ps://doi.o g/ 10.1371%2Fjou nal.pone.0188481. Duda, R.O., Ha , P.E., S o k, D.G., 2000. Pa e n Classi ica ion. Wiley-In e science. Fe dinandy, B., Ge encs´ e , L., Co ie i, L., Pe ez, P., Új ´ a y, D., Csizmadia, G., Mikl´ osi, ´ A., 2020. Challenges o machine lea ning model alida ion using co ela ed beha iou da a: e alua ion o c oss- alida ion s a egies and accu acy measu es. PLoS One 15 (7), e0236092. h ps://doi.o g/10.1371/jou nal.pone.0236092. Ge encs´ e , L., Vasa helyi, G., Nagy, M., Vicsek, T., Miklosi, A., 2013. Iden i ica ion o beha iou in eely mo ing dogs (Canis amilia is) using ine ial senso s. PLoS One 8 (10), e77814. h ps://doi.o g/10.1371/jou nal.pone.0077814. G and View Resea ch, 2018. Pe Wea able Ma ke Size, Sha e & T ends Analysis Repo by Technology, by Applica ion (Iden i ica ion & T acking, Beha io Moni o ing, Facili a ion, Sa e y, Medical Diagnosis), and Segmen Fo ecas s,2018 – 2025. Ma ke Resea ch Repo . Guillo , M., Mo eau, M., Hei , M., Ma el-Pelle ie , J., Pelle ie , J.P., T oncy, E., 2013. Cha ac e iza ion o os eoa h i is in ca s and meloxicam e icacy using objec i e ch onic pain e alua ion ools. Ve . J. 196, 360–367. h ps://doi.o g/10.1016/j. jl.2013.01.009. Halsey, L.G., Shepa d, E.L.C., Wilson, R.P., 2011. Assessing he de elopmen and applica ion o he accele ome y echnique o es ima ing ene gy expendi u e. Compa a. Biochem. Physiol. Pa A 158, 305–314. h ps://doi.o g/10.1016/j. cbpa.2010.09.002. Hamme la, N.Y., Ki kham, R., And as, P., Ploe z, T., 2013. On p ese ing s a is ical cha ac e is ics o accele ome y da a using hei empi ical cumula i e dis ibu ion. P oceedings o he 2013 In e na ional Symposium on Wea able Compu e s 65–68. h ps://doi.o g/10.1145/2493988.2494353. Hansen, B.D., Lascelles, B.D., Keene, B.W., Adams, A.K., Thomson, A.E., 2007. E alua ion o an accele ome e o a -home moni o ing o spon aneous ac i i y in dogs. Am. J. Ve . Res. 68 (5), 468–475. h ps://doi.o g/10.2460/aj .68.5.468. Helm, J., McB ea y, A., Fon aine, S., Mo ison, R., Yam, P., 2016. Use o accele ome y o in es iga e physical ac i i y in dogs ecei ing chemo he apy. J. Small Anim. P ac . 57, 600–609. h ps://doi.o g/10.1111/jsap.12587. Kumpulainen, P., Valldeo iola, A., Somppi, S., T¨ o nq is , H., V¨ a¨ a ¨ aj¨ a, H., Maja an a, P., Su akka, V., Vainio, O., Kujala, M.V., Giza dino a, Y., Vehkaoja, A., 2018. Decembe . Dog ac i i y classi ica ion wi h mo emen senso placed on he colla . P oceedings o he Fi h In e na ional Con e ence on Animal-Compu e In e ac ion 1–6. h ps://doi.o g/10.1145/3295598.3295602. Ladha, C., Ho man, C.L., 2018a. A combined app oach o p edic ing es in dogs using accele ome e s. Senso s 18 (8). h ps://doi.o g/10.3390/s18082649. Ladha, C., Hamme la, N., Hughes, E., Oli ie , P., Ploe z, T., 2013. Dog’s li e: wea able ac i i y ecogni ion o dogs. UbiComp’ 13 P oceedings o he 2013 ACM In e na ional Join Con e ence on Pe asi e and Ubiqui ous Compu ing 415–418. h ps://doi.o g/10.1145/2493432.2493519. Ladha, C., Belshaw, Z., O’Sulli an, J., Ashe , L., 2018b. A s ep in he igh di ec ion: an open-design pedome e algo i hm o dogs. BMC Ve . Res. 14 h ps://doi.o g/ 10.1186/s12917-018-1422-3, 107. Li le, D., Johnson, S., Hash, J., Olson, S.A., Es es, B.T., Mou os, F.T., Lascelles, B.D.X., Guilak, F., 2016. Func ional ou come measu es in a su gical model o hip os eoa h i is in dogs. J. Exp. O hop. 3, 17 h ps://dx.doi.o g/10.1186%2Fs40634- 016-0053-5. Ma in, K.W., Olsen, A.M., Duncan, C.G., Due , F.M., 2016. The me hod o a achmen in luences accele ome e -based ac i i y da a in dogs. BMC Ve . Res. 13 (1) h ps:// doi.o g/10.1186/s12917-017-0971-1. Mo eau, M., Siebe , S., Bue ke , A., Schlech , E., 2009. Use o a i-axial accele ome e o au oma ed eco ding and classi ica ion o goa s’ g azing beha io . Appl. Anim. Beha . Sci. 119 (3–4), 158–170. h ps://doi.o g/10.1016/j.applanim.2009.04.008. Mo ison, R., Reilly, J.J., Penp aze, V., Pendlebu y, E., Yam, P.A., 2014a. 6- mon h obse a ional s udy o changes in objec i ely measu ed physical ac i i y du ing weigh loss in dogs. J. Small Anim. P ac . 55, 566–570. h ps://doi.o g/10.1111/ jsap.12273. Mo ison, R., Penp aze, V., G eening, R., Unde wood, T., Reilly, J.J., Yam, P.S., 2014b. Co ela es o objec i ely measu ed physical ac i i y in dogs. Ve . J. 199 (2), 263–267. h ps://doi.o g/10.1016/j. jl.2013.11.023. Na han, R., Spiegel, O., Fo mann-Roe, S., Ha el, R., Wikelski, M., Ge z, W.M., 2012. Using i-axial accele a ion da a o iden i y beha io al modes o ee- anging animals: gene al concep s and ools illus a ed o g i on ul u es. J. Exp. Biol. 215, 986–996. h ps://doi.o g/10.1242/jeb.058602. Olsen, A.M., E ans, R.B., Due , F.M., 2016. E alua ion o accele ome e in e -de ice a iabili y and colla placemen in dogs. Ve E id. 1 (2) h ps://doi.o g/10.18849/ e. 1i2.40. P es on, T., Bal ze , W., T os , S., 2012. Accele ome e alidi y and placemen o de ec ion o changes in physical ac i i y in dogs unde con olled condi ions on a eadmill. Res. Ve . Sci. 93 (1), 412–416. h ps://doi.o g/10.1016/j. sc.2011.08.005. Ras , W., Kimmig, S.E., Giese, L., Be ge , A., 2020. Machine lea ning goes wild: using da a om cap i e indi iduals o in e wildli e beha iou s. PLoS One 15 (5), e0227317. h ps://doi.o g/10.1371/jou nal.pone.0227317. Robnik-Sikonja, M., Kononenko, I., 2003. Theo e ical and empi ical analysis o Relie F and RRelie F. Mach. Lea n. 53, 23–69. h ps://doi.o g/10.1023/A:1025667309714. Shepa d, E.L., Wilson, R.P., Quin ana, F., Laich, A.G., Liebsch, N., Alba eda, D.A., Halsey, L.G., Gleiss, A., Mo gan, D.T., Mye s, A.E., Newman, C., 2008. Iden i ica ion o animal mo emen pa e ns using i-axial accele ome y. Endange . Species Res. 10, 47–60. h ps://doi.o g/10.3354/es 00084. Vehkaoja, A., Somppi, S., T¨ o nq is , H., Valldeo iola Ca d´ o, A., Kumpulainen, P., V¨ a¨ a ¨ aj¨ a, H., Maja an a, P., Su akka, V., Kujala, M.V., Vainio, O., 2021a. Desc ip ion o Mo emen Senso Da ase o Dog Beha io Classi ica ion submi edDa a in B ie . Vehkaoja, A., Somppi, S., T¨ o nq is , H., Valldeo iola Ca d´ o, A., Kumpulainen, P., V¨ a¨ a ¨ aj¨ a, H., Maja an a, P., Su akka, V., Kujala, M.V., Vainio, O., 2021b. Desc ip ion o mo emen senso da ase o dog beha io classi ica ion. Mendeley Da a. h ps:// doi.o g/10.17632/ xhx934 bn.1 [da ase ]. Wilson, R.P., Whi e, C.R., Quin ana, F., Halsey, L.G., Liebsch, N., Ma in, G.R., Bu le , P. J., 2006. Mo ing owa ds accele a ion o es ima es o ac i i y-speci ic me abolic a e in ee-li ing animals: he case o he co mo an . J. Anim. Ecol. 75, 1081–1090. h ps://doi.o g/10.1111/j.1365-2656.2006.01127.x. Yam, P.S., Penp aze, V., Young, D., Todd, M.S., Cloney, A.D., Hous on-Callaghan, K.A., Reilly, J.J., 2011. Validi y, p ac ical u ili y and eliabili y o Ac iG aph accele ome y o he measu emen o habi ual physical ac i i y in dogs. J. Small Anim. P ac . 52 (2), 86–91. h ps://doi.o g/10.1111/j.1748-5827.2010.01025.x. Yamada, M., Toku iki, M., 2000. Spon aneous ac i i ies measu ed con inuously by an accele ome e in beagle dogs housed in a cage. J. Ve . Med. Sci. 62 (4), 443–447. h ps://doi.o g/10.1292/j ms.62.443. P. Kumpulainen e al.