Topic modelling for routine discovery from egocentric photo-streams
Talavera Martínez, Estefanía; Petkov, Nicolai; Radeva, Petia
Published in:
Pattern recognition
DOI:
10.1016/j.patcog.2020.107330
IMPORTANT NOTE: You are advised to consult the publisher's version (publisher's PDF) if you wish to cite from
it. Please check the document version below.
Document Version
Publisher's PDF, also known as Version of record
Publication date:
2020
Link to publication in University of Groningen/UMCG research database
Citation for published version (APA):
Talavera Martínez, E., Petkov, N., & Radeva, P. (2020). Topic modelling for routine discovery from
egocentric photo-streams. Pattern recognition, 104, [ 107330].
https://doi.org/10.1016/j.patcog.2020.107330
Copyright
Other than for strictly personal use, it is not permitted to download or to forward/distribute the text or part of it without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license (like Creative Commons).
Take-down policy
If you believe that this document breaches copyright please contact us providing details, and we will remove access to the work immediately and investigate your claim.
Downloaded from the University of Groningen/UMCG research database (Pure): http://www.rug.nl/research/portal. For technical reasons the number of authors shown on this cover page is limited to 10 maximum.
ContentslistsavailableatScienceDirect
Pattern
Recognition
journalhomepage:www.elsevier.com/locate/patcog
Topic
modelling
for
routine
discovery
from
egocentric
photo-streams
Estefania
Talavera
a,b,∗,
Carolin
Wuerich
b,
Nicolai
Petkov
a,
Petia
Radeva
ba University of Groningen, Johann Bernoulli Institue, Nijenborgh 9, 9747 AG Groningen, Netherlands
b University of Barcelona, Department Mathematics and Computer Science and Computer Vision Center, Gran Via de les Corts Catalanes, 585, 08007,
Barcelona, Spain
a
r
t
i
c
l
e
i
n
f
o
Article history: Received 17 September 2019 Revised 28 February 2020 Accepted 12 March 2020 Available online 19 March 2020Keywords: Routine Egocentric vision Lifestyle Behaviour analysis Topic modelling
a
b
s
t
r
a
c
t
Developingtoolstounderstandandvisualize lifestyleisofhighinterestwhenaddressing the improve-mentofhabitsandwell-beingofpeople.Routine,definedastheusualthingsthatapersondoesdaily, helpsdescribetheindividuals’lifestyle.Withthispaper,wearethefirstonestoaddressthedevelopment ofnoveltoolsforautomaticdiscoveryofroutinedaysofanindividualfromhis/heregocentricimages.In theproposedmodel,sequencesofimagesarefirstlycharacterizedbysemanticlabelsdetectedby pre-trainedCNNs.Then,thesefeaturesareorganizedintemporal-semanticdocumentstolaterbeembedded intoatopicmodelsspace.Finally,Dynamic-Time-WarpingandSpectral-Clusteringmethodsareusedfor finaldayroutine/non-routine discrimination.Moreover,weintroduce anew EgoRoutine -dataset,a col-lectionof104egocentricdayswithmorethan100.000imagesrecordedby7users.Results showthat routinecanbediscoveredandbehaviouralpatternscanbeobserved.
© 2020TheAuthor(s).PublishedbyElsevierLtd. ThisisanopenaccessarticleundertheCCBYlicense.(http://creativecommons.org/licenses/by/4.0/)
1. Introduction
Withthedynamizationoftheday-by-dayofourcentury,many peopleneedtoimprovethequalityoftheirlife,andthefirststep is to get a better understanding of it. A characterization of the behaviour of a person can help us draw a picture of his or her lifestyle.In[29],theauthorsclaimedthatthedefinitionofpatterns ofbehaviourallowspeople toreachgoalsbycreatingassociations betweenactionsthatarerepeatedinastablecontextandtheir re-sponses.Inourstudy,werelatepatternstothecombinationof el-ementsthatdescribethecontextofthedaysofsomeone,suchas: environment,objectsaroundtheperson,andhis/heractivities. Pat-ternsofbehaviourwerealsodescribedasorderedsequencesof ac-tivities [16]andareimportantelementswhendescribingthe Rou-tine ofa person. At the same time, Routine describes habits and sequencesofactivitiesofsomeone’sdays,andtendstobeunique. More specifically, Routine has been described asregularity in the activity[35].TheabilitytoperformActivitiesofDailyLiving(ADL) directly affectssomeone’s quality oflife. Healthproblems can be detected when certain activities are not performed dueto some issues,such asisolationordepression.Therefore,thediscoveryof
∗ Corresponding author.
E-mail addresses: e.talavera.martinez@rug.nl (E. Talavera), carolin@wuerich.eu (C.
Wuerich), n.petkov@rug.nl (N. Petkov), petia.ivanova@ub.edu (P. Radeva).
the routine of a person is of importance for its later analysis in ordertoassurethehealthylivingofindividuals.
Thecharacterizationofpeople’slifehasbecomean activearea of research with the increasing availability of wearable sensors [25].Lifelogging istheprocess ofcollectingdataaboutthelife of people; this data can describe their activities, emotions and in-teractionsthroughouttheday.In Fig.1,we show aset of photo-streamscollectedbyacamerawearer.Thiscollectionoffersarich sourceofinformationthatallowsunderstandingofthelifestyleof aperson.Morespecifically,byusingwearablecameras,imagescan beautomaticallycollectedfromafirst-person[12],a.k.a.egocentric pointofview ofthecamerawearer.Egocentricimagesarea valu-ablesourceofinformationinmanydomainsduetothesimilarity tohumanperceptionandmemory.However,egocentriccollections useto belarge(oforderofthousandsofpictures perday),which makesdifficultitsanalysis.Inthiswork,werelyonlongtemporal resolution(2fpm)egocentricimagesforthediscoveryandstudyof Routine-relateddaysofpeoplesincetheyallowtomonitorand vi-sualizemostoftheirday.ThediscoveryofRoutineandNon-Routine daysfromegocentric photo-streamsis animportant stepfor sev-eral applications,such as: self-awareness i.e. how doesmy daily lifelooklike?;monitoringpatientsorassistanceofelderlypeople (itis essentialtoknowtheperson’scommonbehaviour and Rou-tine) [9]; or, for memory enhancement and rehabilitation, which benefitsfromstructuring thephoto-streaminto Routineand Non-Routine toeasily find importanteventsused in memory reminis-cencetherapyandinterventions[28].
https://doi.org/10.1016/j.patcog.2020.107330
Fig. 1. Example of images recorded by one of the camera wearers.
Routine-related days have commonpatterns that describe sit-uationsof the daily life of the person.However, Routine hasno concretedefinition,sinceitvariesdependingonthelifestyleofthe individual under study. Therefore,supervised approaches are not usefulduetotheneedforpriorinformationintheformof anno-tated dataorpredefined categories.Forthe discovery of routine-related days, unsupervised methods are necessary to enable an analysisof the datasetwith minimal prior knowledge. Moreover, weneed to apply automatic methods thatcan extract andgroup thedays ofan individual using correlated dailyelements. In this paper,weproposetoapply TopicModelling(TM)technique[5]to detectcorrelatedelementsoftheindividual’sday(e.g.objectsthat appeartogetherofteninthe environmentofthewearer).We use TM asan unsupervised approach forthe analysisof behavioural habitswiththefinalgoalofdetectingRoutinefromegocentric im-ages and thus, to describe and understand the daily patterns of conductofthecamerawearer.Theanalysisoftheappearingtopics throughouttherecordeddaysallowstheunderstandingofthe dif-ferentenvironments wheretheuserspends time:working, shop-ping,walkingoutside,etc.Theseelementsdefinethecontextofthe lifestyleoftheperson.Ourgoalistoaddresstheroutinediscovery byanalyzingtheappearanceofthesepatternsinthelifeofa per-son.Thispatterngiveustheopportunitytocompareandevaluate days.TheyalsoallowustodescribewhatRoutinerepresentsfora persongivenacollectionofhisorherdays.
Inthiswork,weproposetoapplyTMtoourproblemby trans-latingcollected egocentric photo-streams into documents, as we describe in Section 3. We select this technique because it has demonstratedto be a powerful tool forthe discovery of abstract topics appearing in collections of documents, audio, and images [15,17,21,22]. The input images are translated to a Bag-of-Word (BoW)representation,whereanimageisdescribedby theobjects aroundthewearer,activities ofthewearerandthescenethe im-agedepicts.Next,theBoWisconvertedtoanewrepresentationof thedayinterms ofaset ofdiscoveredprobabilistic topics.Then, thefollowingstepistodiscoversimilardays.Routinecanpresent dailysmallvariationsthus,thesimilaritymeasureusetocompare performedactivities duringthedayby thecamerawearer should be tolerantto smalldifferences. Forinstance, having breakfastat 6amandgoing toworkfrom7amto5pmexhibitsthesame Rou-tineashavingbreakfastat7amandworkingfrom9amto7pm.We argue that this allows flexibility in the occurrence of performed activities during the day while temporal order among day ele-ments ismaintained. Therefore,in ourmodel, we define similar-itiesamongdays by evaluatingdistances betweentime-slotsofa
certain duration. To discover similar days we use Dynamic Time Warping for the computationof similarities/distances among the collectedphoto-streams,allowing thatdaily habitsaretolerantto smalldifferencesinstartingtimeandduration.
Thecontributionsofthisworkarethefollowing:
• Weintroduceanautomaticunsupervisedpipelineforthe iden-tification and characterization of Routine-related days from egocentricphoto-streams. Thispipelinecan beadapted to dif-ferentcharacterizationsofdays.Ourmodelisbasedonthe top-icsthatdescribetheday-by-dayfromegocentricphoto-streams fortheirclassificationintoRoutineandNon-Routinedays. • We presenta newegocentric dataset describing the daily life
of the camera wearers. It is composed of a total of 100.000 images,from 104 daysrecorded by 7different users. We call it EgoRoutine and together withits ground-truth are publicly availableinhttp://www.ub.edu/cvub/dataset/.
This paperis organizedas follows:in Section 2, we highlight relevant work related to the routine discovery. In Section 3, we describetheapproachproposedforRoutinediscovery.InSection4, weintroduceourEgoRoutinedataset,outlinetheexperiments per-formedandtheresultsobtained,anddiscusstheachievedresults. Finally, inSections 5and 6, we discussour findings and present ourconclusions,respectively.
2. Relatedworks
Inthissection,we describehowtheroutinebehaviourof peo-plewasstudiedbeforetheraiseofwearabledevicesandwhathas beenstudiedsincethen.
2.1. Routinesfrommanuallyannotateddata
Themanualannotationofdailyhabitstendtobecommon prac-tise forits later analysis by either the own person [3]or physi-cians[39].In[3],manuallyrecordedinformationabouttheability ofsomeoneperformingADLwasexaminedtoclassifythepatients’ dependence,aseitherdependentorindependent.Also,in[39]the authorsstudieddiariesfrom70undergraduatestudents,whorated theassiduityofactivityduringthepreviousmonththrougha ques-tionnaire.
2.2. Automaticroutinediscoveryfromsensorsdata
Withtheincreasingavailabilityofwearablesensors,theaimfor automaticdatacollectingandunderstandingthebehaviourof
peo-plehavebecomeactiveareasofresearch.Thesesensorsallowthe automatic collectionof big amount ofdata describing the life of the person who uses them. One of the first works on analyzing regularities inhuman behaviour from a large scaledataset in an unsupervised mannerwaspresentedin[13].Themodelreliedon informationfrommobile phones,suchaslocations, Bluetooth de-vice proximity, application usage, andphone status.Other works reliedondatacollectedbysensorsplacedinsmarthomes,suchas theonein[26].
One oftheseminal worksonroutine discoverywaspresented in [34]that applied a Latent DirichletAllocation(LDA)model for detecting activities and a subsequent assessment of the similar-ity of a person’s days. There, topic modelling was employed to discoverdailylife activities relatedto rehabilitationpatientsfrom wearable sensors. Specific activitygroups were applied to define the user’s routine. The main 6 categories are eating/leisure (so-cial interactions, eating, playing games), cognitive training (using pc, puzzles), medicalfitness, kitchen work (household activities), motor training, and rest. In [15], the authors focused on Routine discovery by analyzingthe localizationpatterns in a phone loca-tiondatasetcollectedby97peopleover oneyear.Theirproposed modelisbasedonLDAandwordanalyses thatarebuiltbasedon location sequences. Sequences of words are defined by translat-ing the pre-defined locations ‘home’, ‘work’, ‘others’ and ‘no re-ception’ to H, W, O, and N, respectively. Combining a fine-grain (30 minutes) andcoarse-grain (several hours)consideration, they construct a bagrepresentation oflocation sequences. Every loca-tionsequence consistsofthreeconsecutivelocationlabelsforthe fine-grain intervals, followed by a number indicating the coarse-grain time-slot. Thisapproachidentifies Routines whichdominate theentiregroup’sbehavioursuchas‘goingtoworklate’or ‘work-ingnon-stop’.Furthermore,theycharacterizeorclassifyindividuals by thoseRoutines.Fromanotherperspective,in[4],thebehaviour information comes from phone GPS location and is used to as-sess thesimilarity ofaperson’s day.The authorsapplied a mod-ifiedversionofDynamicTimeWarping(DTW)[24]methodto se-quencesofGPSpointssampledatanintervalof10seconds. There-after, a spectral clusteringalgorithm isemployed to cluster simi-lardaysandfindanomalousbehaviours.The authorsin[44] pro-posed amodel forthe discovery ofclustersof dailyactivity rou-tines based on accelerometer data,which describesthe expendi-turedataandsteps.Themodelappliesalow rankandsparse de-composition of thedata signal tolater isolate routine and devia-tions astwo differentsetsofclusters. DTWandhierarchical clus-teringareusedforthecomputationofpairwisedistancesandfinal classification,respectively.
2.3. Routinefromconventionalimages
In [40], the authors addressed the problem of recognition of routinechangesfromshort-termvideosequences.Notethat short-term refers to shortly defined time-slots (e.g. 3–4 hours as it is the case of a GoPro data) while long-term tends to define the continuouscollectionthroughouttheday.Thedatasetin[40] was recorded by a staticcamera atthe entranceof a kitchenand for periods of time in 6 consecutive days, in 3 different years. In their approach, they first proposed to define a model per year. Thismodelrepresentsthestructureofthesequentialactivities per-formed bytheindividual duringthat weekandmakes useof Dy-namicBayesian Networktoestimate the similarityamongsliding windows of the collected video sequences against the evaluated model.Byevaluatingthedifferencesbetweeneachtimeframeand the model,their algorithmdetects thechanges between yearsin the performedactivities when the person is in the kitchen. De-spitetheexcellent resultsofthiswork,thismethodisappliedon strongly controlled environments under the field of view of the
staticcamera andso arenot applicable to detect routine daysof individuals.
Theanalysisofthebehaviourofpeoplehasbeenpreviously ad-dressed for personalizedapplications such asroute planning[7], travel[31,41]orpoint-of-interestRecommendations[23,43],among others.In[41],the authorsuse dynamictopicmodellingto mine the visited places describedby intentionally collected photos by individuals. Based on the discovered topics, other similar loca-tions are recommended. However, none of the above-mentioned approachescoulddescribeanddealwiththeanalysisofcollections of photo-streams, which are what we believe can help to better understandthebehaviouroftheuser.
2.4.Routinefromegocentricimages
The availability of wearable cameras allows to collect large amount of egocentric photo-streams, showing a first-view per-spective ofthe performed activities by the camera wearer. Since theegocentricvisionfieldemerged,severalworkshaveaddressed the analysis of such collections of data from different perspec-tives:activityrecognition[18–20],socialinteractions characteriza-tion[1,2], food-scenesclassification[36],photo-stream segmenta-tion [11], and sentiment analysis [38]. Especially difficult is the problemofanalysisoflong-termegocentricphoto-streams(e.g. ac-tivityrecognition),astheyarerecordedwithalowerframerate(2 fpm)andthereforeprovide sparsercontextual information.Other relatedworks mainly focus on the analysisof ADL. For instance, the workspresented in [14] and[20] analyze egocentric images, focusingonrecognizingtheactivitiesthecamerawearer was per-forming.Thesestudiesdonot godeeperinto theanalysisofhow regularly the recognized activities or environment appear in the recorded photo-streams. Such pattern of appearance is what we believewillallowustodiscoverRoutine-relateddays.
Whereasmostofthelong-termRoutineanalysisapproachesrely on mobile phone locations or sensor data, our approach models patternsofbehaviourbasedonvisualdatafromegocentricimages. Thissourceofdataallowsustounderstandthesurroundingworld andtogiveavisualexplanationtoourfindings.Tothebestofour knowledge, the only work addressing routine behavioural analy-sisfromegocentricimagesis [37],beingaverypreliminary work anda proofofconceptofourproposalhere. There,weaddressed theclassificationofegocentricphoto-streams intoRoutine or Non-RoutinerelateddaysasanAnomalyDetectionproblem.Thatmodel achievedan average of 76% Accuracy and 69% F-Score.However, thereweaddressedtheproblemofRoutinediscoveryfrom egocen-tricphoto-streams followinga very basicandstraightforward so-lution.Theproposed modelwasbasedon theIsolationForest al-gorithmthatpartitionsthedatabasedonananomalyscore.Aday wasdescribedastheaverageoftheobtainedglobalfeaturesforits sequenceofimages.Themethodevaluatesthegivenfeaturevector asthe descriptor fora day. Moreover, there we didnot describe patternsofbehaviourofpeoplesincedayswererepresentedbythe aggregationoftheglobalfeaturesofallimagesthatcomposedthe photo-stream. In contrast withthe mentioned above, this article goesonestepfurtherbyautomaticallydiscoveringroutinesaswell asvisualizing anddescribing behavioural patternsof the camera wearerfromhisorhercollectedphoto-streams.
3. Discoveryofroutine-relateddaysfromegocentric photo-streams
Inthissection, we describe ourproposed model forthe char-acterization of egocentric photo-streams for their later classifica-tionintoRoutineandNon-Routinerelateddays.Fig.2illustratesthe main steps that ourmodel follows givena set ofcollected
long-Fig. 2. Illustration of the proposed pipeline for the discovery of routine from sets of egocentric photo-streams collected by a user. The model proceeds as follows: (a) image semantics extraction, (b) temporal documents construction, (c) topics day representation, and finally, (d) unsupervised routine discovery.
termtemporalresolutionphoto-streams.Below,wedescribein de-tailhowtheyareimplemented.
a)Imagesemanticsextraction
Describing sequences of photo-streams isnot atrivial task due to the unknown visual content. In this work, we pro-posetodescribeourdailyrecordedimagesthroughdetected concepts byan alreadypre-trainedCNN.Forabroad analy-sisofthescenedepictedonagivenimage,wemakeuseof CNNspre-trainedfortherecognitionofobjects[8,30],places [45],andactivities[6].
LetusconsiderthatforeachimageItheCNNsreturn,Lr la-belsrelatedtoatotalofRconceptsfoundintheimages; ob-jects,scene,andactivitiesofthewearer.Thus,eachimageis representedby aBag-of-Wordscomposedofthesedetected semanticconcepts(CNNlabels).
b)Temporaldocumentsconstruction
To model the patterns of behaviour of the camera wearer, we embedthe detected semanticlabelsextractedfromthe egocentric images into a temporaldocument. The detected conceptsbytheCNNsrepresentthewordsthatdescribethe dayi.e.thatformthedocument.
Inordertomaintainthetemporalinformationaboutthe ap-pearanceoftheextractedsemantics,wedefineJtime inter-vals within the day (e.g. from 7-9h, 9-11h, etc.). For each time-interval we estimate the frequency of appearing of eachconcept(Lr,r=1...R).Forthetime-intervalsinwhich no imagesare taken, we createa dummyvariable. Hence, eachdayisrepresentedbyavectorofJ× Rdimension. GivenasetIuofegocentricphoto-streams(days)foruseru, a matrixMi,j isconstructed whereeach ofits elements(ij) corresponds to day i=1,...,
|
Iu|
, and j=1,...,J× R. This temporal document is composed of the concepts detected in the images recorded at a specific range of time. Thus, the proposed modeltranslatesarecorded daythat is com-posedofasequenceofegocentricimages,toatemporal doc-umentrepresentedbythematrixMijdefinedintermsofthe frequency of the detected concepts (words) in the photo-stream.c)Topicsdayrepresentation
Topicmodellingallowsthetransformationofthedatasetby factorisation ofa setD ofdocuments. Adocument is com-posed of a vector of words frequencies, and at the same time, it is assumedthat it definesa certain number, K, of topics. In thiswork, we rely on Latent Dirichlet Allocation (LDA) [5], a topicmodelling approach that is a generative probabilistic modelappliedtoexplain multinomial observa-tions usingunsupervisedlearning. TheLDAmethod follows agenerativeprocessdescribedasfollows[5]:
(a)Choose
θ
i~ Dirichlet(α
),wherei∈{
1,...,D}
. (b)ForeachoftheNiwordswijindocumenti:i.chooseatopiczij ~ Multinomial(
θ
i)ii. choose a word wij from P(wij|zij,
β
) ~ Multinomial probabilityonthetopiczij.where the parameters of the multinomials for topics in a document
θ
i andwordsin atopiczij haveDirichletpriors,Dir(
α
) and Dir(β
) respectively. The probability of a corpus withDdocumentsisdefinedasfollows:P
(
D|
α
,β
)
= |D| i=1 P(
θ
i|
α
)(
Ni j=1 zi j P(
wi j|
zi j,β
)
P(
zi j|
θ
i)
dθ
i wheretheparametersα
andβ
aresampledonlyonceinthe processofgeneratingthecorpus,whilethevariablesθ
i are sampledonceperdocument.Lastly,thevariableszijandwij areword-levelvariableswhicharesampledonceperwordj ineachdocumenti.Asaresult,givenacorpus(set)ofDdocumentsandKtopics tobediscovered,LDAgives[5]:
• thestructureorcombinationofwords thatbestfits the numberoftopics,bygivinga topic-wordmatrixP(wij|zij,
β
) where each element of it defines the probability of assigningwordwij totopiczij.• adocument-topicmatrixP(zij|
θ
i) sothateach elementof itdefinestheprobability ofatopiczij forgivena docu-mentθ
i.In our case, we apply the LDA to decompose the ele-ments Mi,j ofthe temporal documents M corresponding to dayi andtime-slot j.LDA returnsa document-topicmatrix P(zij|Mij)withtheprobabilitiesofallKtopicsassociatedwith each element Mij andthetopic-words matrix P(wij|zij) that defines therelations betweentopics andwords. This is il-lustrated inFig. 3 showinga day represented by themost importanttopics(withthehighestprobability)andthe rela-tionsbetweentopicsandwords.
d)Unsupervisedroutinediscovery
Once we have the representation of each day in terms of the most relevant topics with their probabilities, we need to findsimilaritiesamong daysfor their later classification asRoutineorNon-Routinedays.Forexample,weexpectthat days that used to repeat (e.g. defined by topics related to breakfast,metro,work,lunch,work,metro,anddinner),appear frequentlyandthuscorrespondtoauser’sroutinedays. At thispoint, a day isrepresented asa J-dimensional vec-tor,whereeachelementisaK-dimensionalvectorcomposed oftheprobabilitiesofthedetected topicsdescribingit (see Fig. 3). In order to find similar days, we need a metric to comparetopicsrepresentation.However,it shouldbe toler-ant to small temporal differences, since events during the dayscanbeginandlastdifferently.Tothispurpose,we pro-posetoapplyDTW[24]forcomputingthesimilarityof top-ics representation among days. DTW is an algorithm that computes the optimal alignment between two sequences,
Fig. 3. Illustration of how a photo-stream/document ( Day i ) is described by different proportions of topics throughout the day. We present the winning topic for each time- slot, together with the following N = 2 topics with the higher representation.
Table 1
Total number of recorded days and collected images per user.
User ID 1 2 3 4 5 6 7 Total
Num Days 14 10 16 20 13 18 13 104
Images per day 20,521 9583 21,606 19,152 17,046 16,592 10,957 115,430
where one of them might be stretched or shrunken non-linearly along thetime axis.Given two sequences (or vec-tors)correspondingtotwodayrepresentations,awarppath
(
w1,w2,...,wQ)
isconstructed,whereQisthelengthofthe pathandeveryelement wqisa pair(wq[1], wq[2])that in-dicatesthemappingofelement wq[1]inthefirstsequences to element wq[2] in the second one s. Further, wq[1] andwq[2]havetomonotonicallyincrease.Theoptimalwarp path defines the best correspondence of elements of both sequences represented by the path with minimal distance andiscomputedasfollows:
distDTW
(
s,s)
= Q
r=1
dist
(
swq[1],swq[2])
.In our proposed model, we employ the fastDTW algo-rithm[33],whichisanaccurateapproximation oftheDTW method,buthasalineartimeandspacecomplexity.In con-trast to the standard DTW, the fastDTW algorithm shrinks a timeseriesinto smalleroneswithfewerdata points try-ingtopreserveasmuchinformationabouttheoriginalcurve as possible.Given two sequences describing two days, the fastDTWalgorithmcomputesthedistanceamongthemand gives as output the cost of aligning two days, i.e. their dissimilarity. To compare the topicsrepresentation of each time-slot,weapplyEuclideandistance.
DTW only gives the distance betweenpairs of days. Next, we need to discover clustersof similar days.For that pur-pose, we cannot relyonthe daystopicsrepresentation but onthecomputeddistancesamongpairs.Weapplythe Spec-tralclusteringalgorithm[42]overthecomputedaffinity ma-trixofthedistancesbetweenthedays.Thismethoddoesnot makeassumptionsabouttheglobalstructureofthedata,but bases its decision on local evidence of how likely two el-ements (days)might belong to the same cluster. From the affinity matrix,the algorithm constructs a weightedgraph G=
(
Vn,E,We)
, being Vn the set of nodes, E the set of edges and We the weights of the edges. The global opti-mum is then computedby eigen-decomposition. This clus-tering methodrelieson k-Means forthe final classificationandthus,needsanumberkcofclusterstobedefined,which withoutloss ofgenerality, we setto 2forthe discovery of RoutineandNon-Routinerelateddays.
4. Experimentalframeworkandresults
In this section, we detail a newly introduced EgoRoutine dataset. Then, we describe the metrics used for the evaluation ofthe performed experiments.Next, we depict the experimental setupwith theproposed baseline approaches.Finally,we analyze theobtainedresultsatdifferentstagesoftheproposedpipeline. 4.1. Egoroutine-Anegocentricdatasetforbehaviouranalysis
Inthiswork,weproposeandmakepubliclyavailablethe EgoR-outinedataset1.Thisdatasetiscomposedofrecordeddaysby7
in-dividualswhoworetheNarrativeClipcamera2 fixedtotheirchest
and were asked to record their daily life. EgoRoutine consists of 115.430images,fromatotal of104 recordeddays.InTable1and Fig.4, we indicate the numberof daysand imagescollectedper user. The camera wearerscaptured information abouttheir daily Routine,takingpicturesofthe activitiesthey performedandtheir occurrenceaswellasthepeoplewithwhomtheyinteracted.
GTevaluation:Thecollecteddatasetwaslabelledby6 annota-tors whowere askedto classify daysinto Routine orNon-Routine related.Theannotatorsgotthefollowingdefinition“LifeRoutineis asequence ofactions which are followedregularly, oratspecific intervalsof time, dailyor weekly”. Days were shownto them in theformofamosaics.
InFig.5,wepresentarepresentationofsomeofthecollected photo-streamsofUser1withtheirfinalroutine(R)orNon-Routine (NR) labels given on the right. In Table 2, we present the sum-maryofthelabels givenby thedifferentannotators. Fromthe la-belling results we can deduce that defining what is Routine and Non-Routineisnotaneasy task.Routinecanbeeasilyverbally de-scribed, but it becomes challenging when we want to discover
1http://www.ub.edu/cvub/dataset/ . 2http://getnarrative.com/ .
Fig. 4. Average number and variance of egocentric images per recorded photo-stream for the 7 users. Between parenthesis, we show the number of recorded days per user.
Table 2
Summary of the agreement among the 6 individuals that labelled the collected photo-streams into Routine or Non-Routine related days.
Class Six Agree Five Agree At Least Four Agree At Least Three Agree Total
All 47 29 18 10 104
Routine 35 22 8 0 65
Non-Routine 13 7 9 10 39
itthrough the analysis of sequences of imagesdescribing a long period of time. We observed that in most cases, the annotators agreedwhenlabellingdaysrelatedto Routine.However, the Non-Routinerelateddaysweremoredifficulttoperceiveleadingto dis-agreementamongtheannotators.Forthefinaldistinction,wehave consideredasRoutine related dayswhen morethan 4annotators agreedonthelabel. Incaseofadraw,thedayislabelledas Non-Routine related. Therefore, from a total of 104 recorded days, 65 daysare Routinerelated, and39areNon-Routinerelated.In Fig.6 wepresentthenumberoflabelleddaysperuserintoRoutineand Non-Routine.Ifweextrapolatetoacommonlifescenario,then104 dayscorrespondtoalmost15recordedweeks.Iftheusersfollowed whatcouldbeconsideredascommonRoutine,whereaweekhas5 workingdaysand2weekenddays,in15weekswehave30 week-enddaysand75workingdays.Thiscouldbeanexplanationofthe resultedlabelssinceitisproportionaltotheworkingdaysreported bythecamerawearers.
4.2.Evaluation
Inthissection,wedescribethemetricsthatweusetoevaluate ourproposedmodelforthediscovery ofRoutine andNon-Routine relateddays.
Thediscoveryofroutinebehaviourisanunsupervisedproblem withnon-trivialevaluation.Weevaluatetheresultsintermsof Ac-curacy(A),Precision(P)andRecall(R)andF1scoreintermsofTrue
Positives (TP), TrueNegatives (TN), False Positives(FP), andFalse Negatives(FN),whenclassifyingdaysintoRoutine orNon-Routine, definedasfollows: F1= 2P· R P+R,P= TP TP+FP,R= TP TP+FN,Acc= TP+TN TP+TN+FP+FN Moreover,sincetheproposedpipelineforthediscoveryof rou-tine behavioural patterns is composed of several steps, we also present qualitative results of the intermediate steps of our pro-posal.
4.3. Implementationsetting
Regarding the concepts detected inthe egocentric images, we performanablationstudyusingthefollowingdifferentCNNs:
1. Objectsdetection:DetectedobjectsbyYolo[30]andXception [8].Thesemodels weretrainedon theCOCO[27]and Ima-geNetdataset[10],respectively.
2. Scenerecognition:Werepresentanimagebythetop-1 prob-abilityscenelabelobtainedbytheVGG16,apre-trained net-workpreviouslytrainedonthePlaces365dataset[45]. 3. Activitiesrecognition:Weusetheactivitylabelsgivenbythe
CNNproposedin[6],whichwastrainedfortherecognition of 21 different daily activities. We selectthe activity label withthehighestprobabilityperimage.
Fig. 5. Example of selected images throughout some of the recorded photo-streams of User1. On the right, we can see the given ground-truth (R for routine and NR for non-routine) and the predicted binary label by the best combination of parameters (1 for Non-routine and 0 for Routine days).
Fig. 6. Number of Routine and Non-Routine days for each user (U) in the EgoRoutine dataset.
ConcerningDTW, we usetheEuclidean metrictocompute the distanceamongsamples.Finally,withrespecttotheSpectral clus-tering, we set k equal to 2to discover Routine andNon-Routine relateddays.
4.4. Experimentalsetup
We evaluatethe performanceof thedifferentsteps ofour ap-proach:
• Imagesemanticsextractionintermsofthedetected concepts intheegocentricimagesbythepre-trainedCNNsasdescriptors oftheegocentricphoto-streams.
• Temporal documents construction by the conversion of photo-streamsconceptstodocuments.Toevaluatetheeffectof this,wetestthefollowing:
1. Long duration time-slots: We define J number of time-slots following the ones proposed in [15]: 0am-7am, 7am-9am, 9am-11am, 11am-2pm, 2pm-5pm, 5pm-7pm, 7pm-9pm,9pm-12pm.
2.Shortdurationtime-slots:Ofonehoureach,00:00-01:00, 01:00-02:00,02:00-03:00,etc,witharesultof24 time-slots.
• Topicsdayrepresentation,weevaluatetheimportanceandthe robustnessoftheproposalonthenumberoftopics.Moreover, westudytheneedofindividualvs.generictopicmodelsin or-der to explore if the information about the routine of other users improve the final classification. Given multiple camera users,theLDAmodelcanbecomputedeitherusingtheimages ofall users(generic) orconsidering thesetofdocuments col-lectedbyeachpersonseparately(personalized).
• Unsupervised routine discovery ofphoto-streams. We assess thegoodnessoftheproposedclusteringmethodforthe discov-ery of routine-relateddays, comparing it to the one achieved whenusingtheAgglomerativeHierarchicalClustering[32]forthe discriminationamongdays.
4.5.Resultsanddiscussions
Next,wepresentquantitativeandqualitativeresultsofthe per-formanceon thedifferentstagesofour approachforroutine dis-coveryvalidatedonourEgoRoutinedataset.
• Image semanticsextractionperformance:intermsofthe de-tected concepts:objects,activities andscenes.Withinan abla-tion study we evaluate the performance ofthe different con-cept descriptors when they are considered separately or asa combination. InTable3,wedepict theperformanceofthe ex-periments obtained.As itcanbeobserved,thecombinationof labels of detected objects,activity andplaces better describes the data leading to the best results when addressing routine discovery,withAcc=80%andF1=77%.Thismakessensesince
a richerdescriptionoftheimage helpsto betterdrawthe de-scription of the behaviour of people. Depending on the final goal and application, it could be that independently studying information about activities, objects and/or places helps de-scribebettertheroutineofpeople.
E. Ta la ve ra , C. Wu er ic h and N. Pe tk o v et al. / Pa tt er n R ecognition 10 4 (2020) 1 0 7330
TimeSlot Clustering #Topics Acc F1 P R Acc F1 P R Acc F1 P R Acc F1 P R Acc F1 P R
Personalize Per Hour SpClus 2 0.72 0.68 0.70 0.71 0.71 0.68 0.73 0.75 0.72 0.70 0.72 0.73 0.68 0.65 0.69 0.70 0.72 0.69 0.70 0.72 4 0.75 0.73 0.74 0.77 0.72 0.71 0.74 0.77 0.72 0.69 0.70 0.71 0.78 0.76 0.77 0.81 0.75 0.72 0.74 0.75 6 0.72 0.70 0.73 0.76 0.76 0.73 0.74 0.76 0.76 0.73 0.75 0.77 0.74 0.72 0.75 0.78 0.76 0.72 0.74 0.76 8 0.78 0.75 0.76 0.79 0.76 0.73 0.75 0.78 0.77 0.75 0.78 0.81 0.71 0.70 0.75 0.76 0.77 0.73 0.76 0.80 10 0.73 0.72 0.75 0.78 0.73 0.70 0.72 0.74 0.69 0.66 0.69 0.71 0.72 0.69 0.72 0.74 0.74 0.71 0.74 0.75 HierClus 2 0.68 0.64 0.71 0.71 0.66 0.64 0.73 0.74 0.71 0.69 0.74 0.76 0.71 0.69 0.73 0.74 0.71 0.68 0.76 0.74 4 0.75 0.72 0.77 0.77 0.76 0.74 0.76 0.78 0.71 0.67 0.72 0.72 0.75 0.72 0.76 0.77 0.73 0.69 0.72 0.74 6 0.66 0.60 0.66 0.67 0.76 0.73 0.77 0.79 0.71 0.65 0.71 0.69 0.75 0.71 0.78 0.75 0.70 0.68 0.71 0.74 8 0.79 0.75 0.83 0.79 0.72 0.68 0.71 0.71 0.72 0.66 0.73 0.72 0.77 0.75 0.81 0.82 0.75 0.72 0.78 0.77 10 0.72 0.64 0.69 0.68 0.71 0.63 0.67 0.71 0.67 0.61 0.67 0.69 0.76 0.71 0.71 0.75 0.73 0.66 0.74 0.73 As in [15] SpClus 2 0.69 0.66 0.69 0.71 0.66 0.63 0.67 0.68 0.68 0.66 0.71 0.72 0.68 0.67 0.70 0.72 0.69 0.68 0.71 0.73 4 0.72 0.71 0.74 0.77 0.75 0.72 0.75 0.77 0.74 0.72 0.74 0.77 0.75 0.73 0.77 0.79 0.77 0.75 0.77 0.80 6 0.77 0.75 0.77 0.80 0.71 0.68 0.72 0.74 0.72 0.68 0.70 0.72 0.74 0.71 0.74 0.76 0.80 0.77 0.79 0.82 8 0.70 0.67 0.70 0.72 0.66 0.63 0.70 0.70 0.76 0.72 0.73 0.74 0.76 0.73 0.74 0.77 0.72 0.69 0.72 0.74 10 0.76 0.73 0.74 0.76 0.70 0.66 0.72 0.72 0.75 0.73 0.74 0.76 0.77 0.75 0.77 0.80 0.77 0.75 0.76 0.79 HierClus 2 0.73 0.70 0.72 0.73 0.69 0.67 0.72 0.72 0.69 0.63 0.65 0.67 0.64 0.60 0.67 0.66 0.72 0.63 0.64 0.68 4 0.70 0.68 0.72 0.74 0.70 0.68 0.71 0.74 0.69 0.68 0.72 0.74 0.68 0.65 0.69 0.71 0.74 0.73 0.75 0.77 6 0.73 0.72 0.76 0.79 0.63 0.57 0.64 0.65 0.65 0.56 0.60 0.63 0.71 0.69 0.72 0.74 0.75 0.72 0.75 0.75 8 0.66 0.62 0.70 0.69 0.67 0.62 0.68 0.69 0.71 0.66 0.69 0.70 0.71 0.66 0.70 0.71 0.75 0.70 0.71 0.73 10 0.67 0.59 0.61 0.66 0.72 0.64 0.69 0.69 0.67 0.60 0.68 0.68 0.71 0.69 0.72 0.75 0.73 0.66 0.71 0.71 Generic Per Hour SpClus 2 0.74 0.69 0.70 0.71 0.76 0.74 0.76 0.79 0.79 0.75 0.75 0.77 0.72 0.69 0.70 0.72 0.76 0.72 0.73 0.75 4 0.74 0.70 0.73 0.75 0.78 0.74 0.75 0.78 0.77 0.75 0.78 0.80 0.74 0.72 0.75 0.78 0.77 0.74 0.76 0.77 6 0.76 0.72 0.74 0.76 0.75 0.71 0.73 0.76 0.74 0.73 0.76 0.79 0.76 0.74 0.75 0.78 0.75 0.71 0.73 0.75 8 0.72 0.69 0.72 0.74 0.74 0.71 0.73 0.75 0.73 0.71 0.74 0.76 0.76 0.74 0.76 0.78 0.76 0.72 0.74 0.76 10 0.76 0.72 0.74 0.76 0.75 0.72 0.74 0.76 0.73 0.71 0.72 0.75 0.75 0.73 0.76 0.79 0.74 0.71 0.74 0.75 HierClus 2 0.69 0.65 0.69 0.71 0.67 0.59 0.65 0.65 0.68 0.65 0.71 0.72 0.68 0.65 0.72 0.72 0.67 0.63 0.70 0.70 4 0.75 0.71 0.78 0.76 0.74 0.68 0.70 0.73 0.75 0.72 0.77 0.76 0.67 0.63 0.70 0.69 0.74 0.70 0.72 0.74 6 0.72 0.66 0.67 0.71 0.67 0.63 0.71 0.71 0.73 0.68 0.72 0.75 0.79 0.75 0.81 0.76 0.73 0.70 0.75 0.77 8 0.67 0.63 0.77 0.72 0.69 0.65 0.75 0.73 0.73 0.64 0.65 0.70 0.75 0.70 0.76 0.74 0.76 0.73 0.75 0.78 10 0.68 0.66 0.73 0.75 0.74 0.67 0.70 0.70 0.70 0.63 0.71 0.70 0.73 0.69 0.76 0.73 0.76 0.70 0.77 0.74 As in [15] SpClus 2 0.70 0.68 0.71 0.73 0.71 0.69 0.73 0.74 0.67 0.66 0.68 0.71 0.69 0.66 0.70 0.71 0.69 0.67 0.72 0.73 4 0.69 0.66 0.70 0.72 0.71 0.68 0.73 0.74 0.70 0.67 0.68 0.70 0.73 0.71 0.75 0.77 0.78 0.76 0.78 0.81 6 0.75 0.72 0.74 0.77 0.73 0.71 0.73 0.76 0.69 0.65 0.67 0.68 0.74 0.70 0.72 0.73 0.78 0.76 0.77 0.80 8 0.74 0.71 0.72 0.75 0.69 0.64 0.67 0.68 0.72 0.68 0.70 0.73 0.72 0.70 0.73 0.75 0.75 0.72 0.74 0.76 10 0.72 0.69 0.71 0.74 0.73 0.70 0.74 0.76 0.73 0.70 0.72 0.74 0.76 0.74 0.76 0.79 0.76 0.74 0.76 0.78 HierClus 2 0.73 0.68 0.71 0.73 0.67 0.65 0.70 0.71 0.73 0.70 0.71 0.73 0.70 0.64 0.69 0.70 0.65 0.63 0.70 0.70 4 0.68 0.65 0.68 0.70 0.66 0.64 0.71 0.71 0.64 0.58 0.62 0.63 0.60 0.54 0.64 0.63 0.64 0.59 0.65 0.67 6 0.74 0.67 0.68 0.72 0.69 0.64 0.69 0.70 0.70 0.65 0.73 0.70 0.69 0.63 0.75 0.69 0.72 0.67 0.68 0.73 8 0.69 0.64 0.69 0.70 0.67 0.61 0.64 0.64 0.74 0.70 0.74 0.75 0.69 0.61 0.67 0.65 0.70 0.68 0.75 0.75 10 0.75 0.68 0.73 0.73 0.72 0.66 0.70 0.72 0.71 0.67 0.70 0.70 0.75 0.71 0.77 0.75 0.67 0.61 0.67 0.69
Table 4
Results of the proposed pipeline for the best setting of the parameters: analysing the set of collected photo-streams of User1, seeking for 6 topics to describe the data, with time-slots of long duration, and with spectral clustering as the final classifier.
User 1 User 2 User 3 User 4 User 5 User 6 User 7 Avg Acc 0.79 0.74 0.75 0.90 0.92 0.56 0.92 0.80
F1 0.75 0.70 0.71 0.89 0.92 0.50 0.92 0.77
P 0.75 0.75 0.70 0.89 0.93 0.56 0.94 0.79 R 0.86 0.79 0.75 0.89 0.93 0.60 0.92 0.82
InTable5,weshowconceptsthataredetectedbythedifferent evaluated CNNsina givenphoto-stream.Overall,the detected places by the network getclose enough to reality and there-foreareevaluated.Inthecaseofactivity recognition,andsince thenetworkwastrainedwithegocentricimages,theresultsare moreconsistent.Forthedetectionofobjects,YOLOseemsmore consistent when detecting objects ofthe daily living. We un-derstandthatthisisduetothefactthat theCNNwastrained with80differentcategoriescorrespondingtoCommonObjects inContext(COCO [27]).In contrast,Xceptionmight be ableto recognizeuncommonobjectssinceitwastrainedoverabigger dataset composed of 1000 different categories (the ImageNet [10]).Wecanobservesomeinconsistenciesintheclassesgiven bythenetworktrainedoverPlaces365,suchasfindingthe ‘air-plane cabin’ labelearly in the morning. We explain it by the factthatthisnetworkwasnottrainedwithegocentricpictures. Thechangeofperspectivemodifieshowscenesareunderstood, andlightsin the ceilingofan officeor corridorcan be miss-interpretedasthelightsinthecabinofanairplane.
• Evaluation of the temporal documents construction: We studytheeffectonthediscoveredtopicsforthefinal classifica-tionwhenanalyzingtime-slotsofdifferentduration.Time-slots oflonger durationmight affecttheresultby smoothing activ-ities happening during a short time. In contrast,fine-grained time-slotsmight leadto noise inthe final classification. From the resultsshown inTable 3, we can observe that themodel better performs when the day is described by analyzing the timedivisionproposedin[15].Wededucethattime-slotswith alongerdurationsmooththeactivitiesperformedduringshort periodsoftimewhencomparingdays.Afine-grainedtime-slots withanhourdurationmightincludenoisetothedescriptionof aday.
• Evaluation of the topics day representation performance: Topicmodels discoverabstracttopicswithingivendocuments. Anaturalquestionthatmayariseisthedatausedforthe dis-coveryoftopics:shouldtheybediscoveredfromtheset involv-ingallusersortheyshouldbeextractedforeachuser individu-ally?.Ahypothesisisthatifmoredocumentsaregiven(joining alldata),morerobust topicswillbediscovered,andthus, bet-tertheywillbeabletodescribethebehaviouralpatternsofthe camera wearers.Thus,when learningthe topic-word distribu-tion followingthe generic approach, we could take advantage ofa biggerdataset.Anegativeaspectofseekinggeneralization isthat user-specificactivities canbe missed,since theywould becomenotrelevanttobedetected.Incontrast,weassumethat individuallylearnedtopicsmightfindmorepersonalized repre-sentationsofeveryspecificactivityoftheuser,sincetheplaces oftheir dailylife, e.g. theoffice desk orlivingroom of differ-entpeople,mightbe describeddifferently. Therefore,we eval-uate the performance of the model when obtaining the top-icsjustbased onthe collectedphoto-streamsby the user un-der study (personalized approach), orwhen analyzing all the collected photo-streams that compose the EgoRoutine dataset (genericapproach).Fromtheresultsandforthegoalofroutine discovery,thepersonalizedapproachallows themodelto
bet-ter distinguish Routine-related days witha 80% accuracy and 77%F1(seeTable3).
The goodness ofthe modelwhen varyingthe numberof top-ics is also tested. We present results when discovering 2, 4, 6, 8and10 topics.As itcan be observed,the performance of theclassifierishighestwhendiscovering6andaddressing the time-divisionproposed in[15].However, itcouldbethat fora moredetailedanalysisofwhatishappeningataspecifictime, a highernumber of fine-grained time-slots mightdescribe in moredetail,intermsofobjects,activitiesandplaces.
• Evaluation of the Unsupervised routine discovery perfor-mance:WecomparetheperformanceoftheproposedSpectral Clusteringalgorithmwiththeresultsobtainedbythe Agglomer-ativeHierarchicalClustering[32](HC)whenclassifyinginto Rou-tineorNon-Routinerelateddays.HCmethodfollowsa bottom-up approach whereeach data point startsas a singlecluster, andpairsofsamplesarerecursivelymergedfollowingthepath thatminimallyincreasesthegivenlinkagedistance.Theprocess continuesassamplesareclusteredmovingupinthesimilarity hierarchy.We selecttheHC sinceweneedtocompareagainst methodsthatareabletoanalysepre-computeddistance matri-ces.
WecanobserveinTable3thattheSpectralClusteringclassifier leadstoamoreaccuratediscoveryoftheRoutine-relateddays, outperforming theclassification by the HC. Webelieve this is duetotheabilityoftheSpectralclusteringtoadapttocomplex shapesofthedatainthedataspace.
Fora moredetailedunderstandingoftheperformance atuser level,inTable4weshowresultsofthebestperformingmodel. We can observe that for some of the users the classification into RoutineandNon-Routinerelateddays isratherclear,such as forUser 5 or User 7,while for User 6 the classification is closetorandom.Thisisduetodifferencesbetweenthelifestyle of the users. Some of them have a clear distribution of rou-tine(e.g.work)andnon-routine(e.g.non-work)related activi-ties, whileothersrecorded daysforperiodswhentheir activi-tieswerenotfollowinganestablishedroutinepattern.
InFig.5,wepresentsomecollecteddaysofUser1andthe pre-dictedlabelbythebestcombinationofparameters(personalize analysisofdocuments,combinationoflabelsasimages descrip-tors, 6topics,andSpectral clustering).Days predictedas Non-Routinerelatedareassignedlabel‘1’andRoutine-related days label‘0’. Day1ismiss-classifiedasNon-Routine related.From observing thedata, wecan guessthat thisusertends to start workingatnoon anduntillate inthe evening.In contrast,on Day1,User1spentmuchfewerhoursatworkandleftthe of-ficemuchearlier.Thiscouldbeacauseofmiss-classificationby the model.Non-Routine related days contained events where theuserworkedforshortperiodsandspentlongertime inter-acting withcolleagues orfriends. Day 7 isan example where User1wentfordinnertoarestaurantrightafterworkingfora shorttime.
• Final routine characterization and visualization for be-haviourmodelling: Thecharacterizationofdaysbasedon de-tectedconceptsandthelaterinferredtopicshavedemonstrated
Places [45] airplane cabin
90 airplane cabin 167 conference room 49 office 41 airplane cabin 31 reception 28
atrium/public
8 office 113 office 43 airplane cabin 26 bowling alley 14 airplane cabin 26 office
cubicles
8 office cubicles 42 reception 37 computer room 23 airport terminal 10 hotel room 14 Activity [6] WalkingIn 50 Mobile 227 Mobile 60 Working 78 Mobile 30 Talking 50 Shopping 40 Shopping 94 Talking 46 Mobile 39 Driving 25 WalkingOut 37 WalkingOut
36 Working 75 meeting 46 WalkingOut 32 WalkingOut 16 Mobile 27 Yolo [30] person 146 tvmonitor 383 person 202 person 132 person 107 person 198
laptop 38 cup 354 laptop 112 tvmonitor 122 chair 32 chair 155 chair 38 laptop 334 chair 108 keyboard 73 cell phone 23 diningtable 53
Fig. 7. Example of given photo-streams, sample images at several time-slots, their representative topics, and the concepts that compose them. We present results with the following combination of the parameters of our model: activity labels, time-slots as in [15] , 8 topics and personalized approach.
tobearichtoolforbehaviourvisualization.InFig.7wepresent how the found topics could be analysed by the wearer oran expert. As an exampleof visualization, results are shown fol-lowinga personalizedanalysisofthe datacollectedbyUser 1 describedwithactivitylabels,anddiscovering8 topics.Aswe
canobserve,Non-routinerelateddaysdifferfromthe Routine-relateddaysasthefirstonepresentsTopic0andTopic7,which arecomposed ofactivitylabelsdescribingsocial interactionin food-relatedenvironments.Routine-relateddaysaremainly de-scribedbyTopic1,3,4,and5,whichdescribeworking
environ-Fig. 8. Affinity matrix obtained from the distances computed by DTW for the later discrimination as Routine or Non-Routine related days by Spectral Clustering of collected days by users 3 and 7. Days are divided with orange and blue boxes as the two final clusters. On the right, we indicate the ground-truth labels per day. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)
Fig. 9. Visualization using multi-dimensional scaling (MDS) of the distribution of samples for users 1 and 7. Each dot corresponds to a collected day by the user. We use two colors to distinguish between the two classes. The inside color of the dots is the given ground-truth and the colour of the boundaries of the dots represents the classification label (‘R’ black and ‘NR’ red). (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)
ments.Weunderstandthat activitylabelssuch asmobile, talk-ing, and walking Indoor/Outdoor can be understood as screen, meeting,andcommuting,respectively.
To get insight at the classification level, we presentin Fig. 8 theaffinitymatrixthattheSpectralClusteringusesforthe dis-criminationamongthecollecteddaysbyUser3andUser7.The given labelsfor the collecteddays are indicated in the figure on theright ofthe matrix,where ‘R’ corresponds to Routine-relatedand‘NR’toNon-Routinerelated.Inthepresented affin-itymatrix,wehighlightthetwofinal clusterswithorangeand blue.We can observehow inthe caseofthese usersclear R-relatedclustersaredefined,whileNR-relatedclustersare scat-tered.TheaccuracyforUser3andUser7isof75%and92%, re-spectively,whichagreewiththevisualassociationinFig.8 be-tweensimilardaysandgivenlabels.
Furthermore, in Fig. 9 we visually illustrate the produced re-sults of our model for users 1 and 7. We applied Multi-DimensionalScaling(MDS)forthisvisualizationsinceitallows tovisualizespatialdistributionofdatafromtheirsimilarity ma-trixinsteadofexplicitcoordinates/representations.Weuseitto display themutualspatialdistribution oftheuserdays repre-sentations expressed by the obtained similarity matrix when applying the DTW. We can see the ground-truth indicated as the inside color of the sampleand the classification label as theboundariesofthecircles.Inbothcasesblackcorrespondsto Routine-related daysandredtoNon-routinerelateddays.This visualizationallowsustobetterexploreclassificationresults. Moreover, inFig. 10 we can observe the computed silhouette scores for the obtained final routine and non-routine related clusters. Note that the silhouette score can take a value be-tween 1 and-1.Valuescloseto 1indicate differentiable
clus-Fig. 10. Silhouette score per user for the two discovered clusters, Routine and Non- routine related days.
Table 6
Comparison between our previous work introduced in [37] and the model here proposed for routine discovery from egocentric photo-streams.
Method Number of Users Acc F1
Routine discovery [37] 5 0.76 0.69 Routine discovery propose here 0.82 0.79
Routine discovery propose here 7 0.81 0.80
ters, 0 overlapping of clusters, and negative samples repre-sent the wrongclassification ofsamples. We can seehow for the majority ofthe users, routine-relateddays share a higher scorethanthenon-routinerelateddaysclusters.Thisreinforces our hypothesis that routine related days correspond to more compact clusters,while non-routine relateddaysforma more sparsecluster. Intwoofthecases,thesilhouettescoreforthe routinerelateddaysislowerthanfornon-routinerelatedones. Lookingcloseratthedataweobservedthatinthesecasesthere weremorethanoneroutinegroupsofdays.However,the prob-lemofdiscoveringtheoptimalnumberofclustersandthusthe routinesisoutofthescopeofthispaper.
Finally,inTable6we comparetheobtainedresultsforroutine discovery to theroutine discovery in[37].As one cansee the method in[37] run on 5users achieved0.76 ofaccuracy and 0.69ofF1scorewhilethemethodproposedhereachieved0.81
ofaccuracyand0.80ofF1 score.Apossibleexplanationisthat
the workproposed in [37]reliedon theaggregation ofglobal featuresofalltheimagescomposingadayforitsdescription.In contrast,themodelproposed herereliesonsemanticconcepts combined with topicmodelling, DTW and spectral clustering, whichresultsalsoallowunderstandingofwhatishappeningin thelife ofthecamerauser.We alsopresenttheresultsofour methodforthesubsetoffiveusersthatwereanalyzedin[37], withaperformanceofAcc=0.82andF1=0.79.Aswecan
ob-serve,theresultsarequitesimilar:moreover,higher classifica-tionperformance isachievedwhentopicsmodellingDTWand spectral clusteringare applied to the collection ofdocuments composedofdetectedsemanticconcepts.
5. Discussions
In this work, we presented a newmethod forthe analysis of routinebehaviouralpatternsfromcollectedegocentricvisualdata. Wedemonstratedthattheseimagesarearich sourceof informa-tionandthatdetectedconceptsfromtheimagescanhelpusdraw apictureofthelifestyleofthecamerawearer.
picts. Thisis treatedasa document forthe discovery ofabstract topicsdescribing thethemesofthelifestyle oftheindividual un-derstudy.Documentsarefed toanLDAmodelthat organizes se-manticlabelsintotopicscomputingatopic-worddistributionand a document-topic distribution, thus, obtaining topics distribution for each given document.Moreover, we show that using tempo-raldocumentsbasedontime-slotsintowhichdaysaredivided, al-lowsflexibility when comparing the behaviour at differenttimes oftheday. Thedistances betweenthedayscan be computed us-ing DTWtofinally clusterdaysandassignthem intoRoutineand Non-RoutineonesbyapplyingSpectralclustering.
Moreover, we introduced a newEgoRoutine dataset, on which wetestedandvalidatedourproposed model.The datasetis com-posed ofa total of 104 days,recorded by 7 users,and we make itpubliclyavailable3 forthefuturedevelopmentofthislineof
re-search. Theanalysisofthe modelcouldbe improvedby the aug-mentation of the dataset. For further steps in this direction, we need richer data. However, this is not a trivial task and we are workingon it.Moreover, more accurate detected concepts would beofhelpwhendescribingthecollecteddays.Forthis,wewould needtrainednetworksonegocentricimages.
We hypothesize that Routine-related days will share similar traitsandthus,willrepresentacluster.Commonly,Non-routine re-lateddays,tendtobetheonesnon-workrelated.Thesedaysshare theirownroutine-patterns,i.e.therecanbemorethanoneroutine inthe lifeofpeople; cleaning, cooking,orgoing out withfriends coulddescribeone ofthem.Alimitationofourworkisthat Non-Routinerelated days mightnot define a cluster. In futureworks, we plan to evaluate if the combinationof outlier detectionwith topic modelling allows a better understanding of the lifestyle of thecamerawearer.
We hope that our proposed dataset and the shown results will be a call for other researchers who aim to study people’s behaviour for its understanding and providing tools for lifestyle improvement.
6. Conclusions
Inthiswork,weconcludethatbehaviouralanalysisfromvisual datais possible.Moreover, topicmodelsproved to be apowerful tool for thediscovery ofpatterns when addressing Bag-of-Words representationofphoto-streams.Fromtheobtainedresults,we ob-served that discovered topicmodelsfollowing a personalized ap-proachimprovetheclassificationofdays.Thisprovidesamore de-tailedexplanationofwearerdailybehaviour.However,agenericor personalizedapproachcanbe applieddependingonifthegoalis to detectgeneralinformationor peculiaritiesofthe life ofa per-son.Oneoftheimportantadvantagesofthisworkisthe unsuper-vised discovery of routine and non-routine related days.Given a newuser,wecandiscriminateroutinedaysandcharacterizetheir collectedphoto-streams.
Further works will explore the inclusion of outlier detection techniques andthe discovery of specific behaviours,such as: so-cialinteractionsandnutritionalbehaviourbystudyingthe appear-ance of people in certain situations and food-related scenes,
spectively. Furthermore, we are interested in studying how topic modellingandCNNscanbeinterconnected.
Wehopethat ourproposeddatasetandtheshownresultswill beacallforotherresearcherswhoaimtostudypeople’sbehaviour foritsunderstandingandprovidingtoolsforlifestyleimprovement. DeclarationofCompetingInterest
Theauthorsdeclarethattheyhavenoknowncompeting finan-cialinterestsorpersonalrelationshipsthatcouldhaveappearedto influencetheworkreportedinthispaper.
Acknowledgment
ThisworkwaspartiallyfoundedbyprojectsTIN2015-66951-C2, RTI2018-095232-B-C2,SGR1742,CERCA,NestoreHorizon2020 SC1-PM-15-2017 (n 769643), Validithi EIT Health Program and ICREA Academia2014.Thefoundershadnoroleinthestudydesign,data collection,analysis,andpreparationofthemanuscript.Theauthors gratefullyacknowledgethesupportofNVIDIACorporationwiththe donation ofseveralTitanXp GPUusedforthisresearch. The col-lected dataaspartofthestudyandgivenlabelsispublicly avail-able fromthewebsiteofour researchgroup:http://www.ub.edu/ cvub/dataset/
References
[1] M. Aghaei , M. Dimiccoli , C.C. Ferrer , P. Radeva , Towards social pattern charac- terization in egocentric photo-streams, Comput. Vision Image Understanding (2018) 104–117 .
[2] S. Alletto , G. Serra , S. Calderara , R. Cucchiara , Understanding social relation- ships in egocentric vision, Pattern Recognit. 48 (12) (2015) 4082–4096 .
[3] C.K. Andersen , K.U. Wittrup-Jensen , A. Lolk , K. Andersen , P. Kragh-Sørensen , Ability to perform activities of daily living is the main factor affecting quality of life in patients with dementia, Health Qual. Life Outcomes 2 (1) (2004) 52 .
[4] J. Biagioni , J. Krumm , Days of our lives: assessing day similarity from location traces, International Conference on User Modeling, Adaptation, and Personal- ization (2013) 89–101 .
[5] D.M. Blei , A.Y. Ng , M.I. Jordan , Latent dirichlet allocation, Journal of machine Learning research 3 (Jan) (2003) 993–1022 .
[6] A. Cartas , J. Marín , P. Radeva , M. Dimiccoli , Batch-based activity recognition from egocentric photo-streams revisited, Pattern Analysis and Applications 21 (4) (2018) 953–965 .
[7] D. Chen , D. Kim , L. Xie , M. Shin , A.K. Menon , C.S. Ong , I. Avazpour , J. Grundy , Pathrec: Visual analysis of travel route recommendations, in: Proceedings of the Eleventh ACM Conference on Recommender Systems, 2017, pp. 364–365 .
[8] F. Chollet , Xception: deep learning with depthwise separable convolu- tions, IEEE Conference on Computer Vision and Pattern Recognition (2017) 1800–1807 .
[9] P.-C. Chung , C.-D. Liu , A daily behavior enabled hidden markov model for hu- man behavior understanding, Pattern Recognit 41 (5) (2008) 1572–1580 .
[10] J. Deng , W. Dong , R. Socher , L.-J. Li , K. Li , L. Fei-Fei , Imagenet: a large-scale hierarchical image database, IEEE Conference on Computer Vision and Pattern Recognition (2009) 248–255 .
[11] M. Dimiccoli , M. Bolaños , E. Talavera , M. Aghaei , S.G. Nikolov , P. Radeva , Sr– clustering: semantic regularized clustering for egocentric photo streams seg- mentation, Comput. Vision Image Understanding 155 (2017) 55–69 .
[12] A.R. Doherty , S.E. Hodges , A.C. King , A.F. Smeaton , E. Berry , C.J. Moulin , S. Lind- ley , P. Kelly , C. Foster , Wearable cameras in health: the state of the art and future possibilities, Am. J. Prev. Med. 44 (3) (2013) 320–323 .
[13] N. Eagle , A. Pentland , Reality mining: sensing complex social systems, Personal Ubiquitous Comput. 10 (4) (2006) 255–268 .
[14] M. Ermes , J. PÄrkkÄ, J. MÄntyjÄrvi , I. Korhonen , Detection of daily activities and sports with wearable sensors in controlled and uncontrolled conditions, IEEE Trans. Inf. Technol. Biomed. 12 (1) (2008) 20–26 .
[15] K. Farrahi , D. Gatica-Perez , Discovering routines from large-scale human loca- tions using probabilistic topic models, ACM Trans. Intell. Syst. Technol. 2 (1) (2011) 3 .
[16] I. Fatima , M. Fahim , Y.-K. Lee , S. Lee , A unified framework for activity recogni- tion-based behavior analysis and action prediction in smart homes, Sensors 13 (2) (2013) 2682–2699 .
[17] R. Fernandez-Beltran , F. Pla , Latent topics-based relevance feedback for video retrieval, Pattern Recognit. 51 (2016) 72–84 .
[18] A. Furnari , G. Farinella , S. Battiato , Recognizing personal locations from ego- centric videos, IEEE Trans Hum Mach Syst 47 (1) (2017) 1–13 .
[19] A. Furnari , G.M. Farinella , S. Battiato , Recognizing personal contexts from ego- centric images, IEEE International Conference on Computer Vision Workshop (2015) 393–401 .
[20] A. Furnari , G.M. Farinella , S. Battiato , Temporal segmentation of egocentric videos to highlight personal locations of interest, European Conference on Computer Vision (2016) 474–489 .
[21] S. Hou , L. Chen , D. Tao , S. Zhou , W. Liu , Y. Zheng , Multi-layer multi-view topic model for classifying advertising video, Pattern Recognit. 68 (2017) 66–81 .
[22] P. Hu , W. Liu , W. Jiang , Z. Yang , Latent topic model for audio retrieval, Pattern Recognit. 47 (3) (2014) 1138–1143 .
[23] S. Jiang , X. Qian , J. Shen , Y. Fu , T. Mei ,Author topic model-based collaborative filtering for personalized poi recommendations, IEEE Trans. Multimedia 17 (6) (2015) 907–918 .
[24] E.J. Keogh , M.J. Pazzani , Derivative dynamic time warping, SIAM international conference on data mining (2001) 1–11 .
[25] O.D. Lara , M.A. Labrador , A survey on human activity recognition using wearable sensors, IEEE Communications Surveys & Tutorials 15 (3) (2012) 1192–1209 .
[26] C. Li , W.K. Cheung , J. Liu , Elderly mobility and daily routine analysis based on behavior-aware flow graph modeling, International Conference on Healthcare Informatics (2015) 427–436 .
[27] T.-Y. Lin , M. Maire , S. Belongie , J. Hays , P. Perona , D. Ramanan , P. Dollár , C.L. Zit- nick , Microsoft coco: common objects in context, European Conference on Computer Vision (2014) 740–755 .
[28] G. Oliveira-Barra , M. Bolaños , E. Talavera , A. Dueñas , O. Gelonch , M. Garolera , Serious games application for memory training using egocentric images, Inter- national Conference on Image Analysis and Processing (2017) 120–130 .
[29] Society for Personality , Social Psychology , How we form habits, change existing ones, ScienceDaily (2014) .
[30] J. Redmon , A. Farhadi , Yolov3: an incremental improvement, arXiv (2018) .
[31] S. Renjith , A. Sreekumar , M. Jathavedan , An extensive study on the evolution of context-aware personalized travel recommender systems, Inf. Process. Manag. 57 (1) (2020) 102078 .
[32] L. Rokach , O. Maimon , Clustering methods, Data mining and knowledge dis- covery handbook (2005) 321–352 .
[33] S. Salvador , P. Chan , Toward accurate dynamic time warping in linear time and space, Intell. Data Analalysis 11 (5) (2007) 561–580 .
[34] J. Seiter , A. Derungs , C. Schuster-Amft , O. Amft , G. Tröster , Daily life activity routine discovery in hemiparetic rehabilitation patients using topic models, Methods Inf. Med. 54 (3) (2015) 248–255 .
[35] A. Sevtsuk , C. Ratti , Does urban mobility have a daily routine? learning from the aggregate data of mobile networks, J. Urban Technol. 1 (17) (2010) 41–60 .
[36] E. Talavera , M. Leyva-Vallina , M. Sarker , D. Puig , N. Petkov , P. Radeva , Hierar- chical approach to classify food scenes in egocentric photo-streams, J. Biomed. and Health Informatics (2019) .
[37] E. Talavera , N. Petkov , P. Radeva , Unsupervised routine discovery in egocentric photo-streams, 18th Conference on Computer Analysis of Images and Patterns (2019) .
[38] E. Talavera , N. Strisciuglio , N. Petkov , P. Radeva ,Sentiment recognition in ego- centric photostreams, Iberian Conference on Pattern Recognition and Image Analysis (2017) 471–479 .
[39] W. Wood , J. Quinn , D. Kashy , Habits in everyday life: thought, emotion, and action, J. Pers. Soc. Psychol. 83 (6) (2002) 1281–1297 .
[40] Y. Xu , D. Damen , Human routine change detection using bayesian modelling, International Conference on Pattern Recognition (2018) 1833–1838 .
[41] Z. Xu , L. Chen , Y. Dai , G. Chen , A dynamic topic model and matrix factor- ization-based travel recommendation method exploiting ubiquitous data, IEEE Trans. Multimedia 19 (8) (2017) 1933–1945 .
[42] S.X. Yu , J. Shi , Multiclass spectral clustering, IEEE International Conference on Computer Vision 2 (2003) .
[43] Z. Yu , H. Xu , Z. Yang , B. Guo , Personalized travel package with multi– point-of-interest recommendation based on crowdsourced user footprints, IEEE Trans Hum Mach Syst 46 (1) (2015) 151–158 .
[44] O. Yürüten , J. Zhang , P. Pu , Decomposing activities of daily living to discover routine clusters, Conference on Artificial Intelligence (2014) .
[45] B. Zhou , A. Lapedriza , A. Khosla , A. Oliva , A. Torralba , Places: a 10 million image database for scene recognition, IEEE Trans. Pattern Anal. Mach. Intell. (2017) .
Estefania Talavera received her BSc degree in Electronic engineering from Balearic Islands University in 2012 and her MSc degree in Biomedical Engineering from Polytech- nic University of Catalonia in 2014. She is currently a PhD student at the University of Barcelona and University of Groningen. Her research interests are lifelogging and health applications.
Carolin Wuerich received her BEng degree in Electrical Engineering from the Baden-Wuerttemberg Cooperative State University Stuttgart (Germany) in 2017 and her MSc degree in Artificial Intelligence from the Polytechnic University of Catalonia in 2019.
