modeling pedestrian and crowd behaviour: the case of the crystals project

Adaptive pedestrian behaviour for the preservation of group cohesion: observations and simulations

Modeling Pedestrian and Crowd Behaviour: the case of the Crystals Project

Giuseppe Vizzari1,21Complex Systems and Artificial Intelligence Research Center University of Milano-Bicocca2Crystals Project, Center of Research Excellence in Hajj and Omrah (Hajjcore) Umm Al-Qura University, Makkah, Saudi Arabia

OutlineThe context of application: the Hajj, the Mashaer line and the Arafat I stationGroups, as a crowd management concept and a natural, pervasive presence in pedestrian populationGroups in the relevant literatureIn vitro, in vivo, in silico: the Crystals Project approachObservations about groups (in vitro and in vivo)Modeling and simulation (in silico)Results in the Arafat I scenarioConclusions and future developments


The Hajj in BriefAnnual pilgrimage to Makkah, Saudi ArabiaFifth pillar of Islam, a religious duty that must be carried out at least once in their lifetime by every able-bodied Muslim who can afford to do soOver 2,5 millions of people coming from over 150 countriesA precise and articulated system of rituals implying the mass movement of pilgrims over several sites that in some cases are about 20 km distant

The Mashaer LineFive proposed rail lines connecting the holy sites with one another and with MakkahThe southern rail includes 9 stations: 3 in Mina, 3 in Muzdalifah and 3 in Arafat, to replace 35,000 cars and buses and access the Haram and Makkah Central AreaFuture lines to the Holy HaramExtend the southern rail line to Jeddah Airport, with an elevated alignment above the Jeddah Expressway over an 80 Km length

Observations at the Hajj


Observations at the Hajj - Groups as a crowd management organizational instrument

Pilgrims are subdivided into groups of 250 persons following a leader in their movement from the nearby tents area to the platformThe waiting boxes act as waiting areas hosting groups waiting to use ramps or elevatorsThe platform can safely host even more than 3000 pilgrims (the capacity of a train), but the process is aimed at avoiding overcrowding of the platform

Observations at the Hajj - ConsiderationsGroups are used as an organizational instrument to manage crowdGroup arrival is planned, scheduledLeaders decide when and where to move, collaborating with station officersTheir size is relatively large, their cohesion is not extreme...... but inside them smaller sub-groups can be identified and they can be much more compactGroups have different intermediate movement targets, although the same final goal

Group influence in general - ConsiderationsThe presence of groups is pervasive in many events involving large crowdsGroups are simply out there...... its not a matter of deciding if theyre good or bad for the pedestrian flow... its a matter of understanding their impact, in different relevant conditionsThe presence of groups should be carefully considered:Design choices might make it difficult for a group to preserve its cohesion, which is particularly significant in certain situations (e.g. kids, elderly, mobility impaired persons)...... and this would cause stress in group members and congestions, delays in the whole system


Groups in the literature - ObservationsAt least two studies report observations about groupsWillisA, GjersoeN, HavardC, KerridgeJ, KuklaR, 2004, "Human movement behaviour in urban spaces: implications for the design and modelling of effective pedestrian environments" Environment and Planning B: Planning and Design 31(6) 805828Michael Schultz, Christian Schulz, and Hartmut Fricke. Passenger Dynamics at Airport Terminal Environment, Pedestrian and Evacuation Dynamics 2008, Springer-Verlag, 2010Observations carried out in low density conditionsGroups of small size were most frequently observed

Groups in the literature - Modeling and SimulationExtensions to the social force modelHelbing, Theraulaz et al. 2009, 2010Small groups (2,3,4), unstructuredLow to moderate densitiesValidation based on actual observationsXu and Duh, 2010Only couples (groups of 2 pedestrians)Low to moderate densitiesShallow validation based on literature (Daamen, 2004)CA modelsSarmady, Haron, Zawawi Hj, 2009Leaders and followersGroups of 2 to 6 members experimentedNot validatedAgent-based modelsQiu and Hu 2010Structured groups (intra and inter group matrices)Large groups experimented (60 pedestrians)Not validatedGroup members tend to stay close to other group members (additional behavioural component)


The Crystals Project Approach

In silicoIn vitroIn vivo


Experiments in TokyoExperiments carried out by the Research Center on Advanced Science and Technology of The University of TokyoAimed at evaluating the impact of the presence of groups in experimental situationsSpecifically their impact on the formation of lanes and total travel times in relatively high density situationsResults still not published...... However, we can already say that more experiments and observations are needed to draw conclusionsThe influence of groups is not trivial

Admission testUniversity of Milano-BicoccaAdmission test of the Faculty of Psychology at the University of Milano-Bicocca - September 1, 2011Counting activity supported by video footages of the eventAbout two thousand students attended the testAbout 34% individuals, 50% couples, 13% triples and 3% groups of 4 members (!)Statistically validated relationship between group size and velocityAdditional quantitative analyses about the arrival and entrance process, LOSQualitative analysis of group shapes and related phenomena

Vittorio Emanuele II Gallery, MilanPopular commercial-touristic walkway in Milans city centreGoals of the survey:level of density and walkway level of service (A and B);presence of groups (over 84%);group size and proxemics spatial patterns, trajectories and walking speed (groups are slower but their trajectories are shorter);group proxemics dispersion (they preserve cohesion, even if large ones occupy more space)still hard to evaluate spatial arrangement of group members

Group dispersionCouplesTriples4 MembersDistanceCentroid0.58 m(sd 0,22)0.76 m(sd 0,11)0.67(sd 0.12)


A model considering groupsBased on the floor-field CA approach, with significant difference on movement choiceEmploying traditional factors for movement destination choiceGoal orientationPresence of obstaclesPresence of other pedestrians (basic proxemics)A notion of group has been introducedTo generate a generalised effect of cohesion among members of groups... able to overcome goal orientation for certain types of groups (e.g. families, close friends)Speed heterogeneity also introduced (poster on Monday afternoon)

A few formal details Stochastic choice of destination cell; for each cell c, the probability of choosing an action a leading to it is

The utility value of the cell is defines as follows:

whereGoal is associated to the static floor field and Obs to the wall potentialSep is associated to the proxemic repulsionD is an inertia factorOver regulates the possibility of having two pedestrians sharing the same cell in case of high densityCoh and Inter represent group cohesion factors respectively for small simple groups and large potentially structured groups

OverlappingOverlapping is a transient situation in which pedestrians share the same cell... it can sometimes be observed in counterflow situations in which there is not enough space for avoidance It can only happen if local density exceeds a given thresholdThe choice is still penalised (Over 0)No more than two pedestrians can share a single cell[Kretz et al., 2006]

Simple and structured groupsSimple groups are made up of family members, friends, people that know each otherThey often adapt their behaviour to preserve the cohesion of the groupLarge groups can include perfect strangers that share for some time a common goalMembers of this group have a tendency to stay close to each other...... but this tendency is not so strong to prevent group fragmentationAnd generally they are actually structured (they can include other - often simple - groups), so we call them structured

Multipliers of the different components of movement utility are adjusted according to the state of the group

The dispersion of the group causes an increased impact of simple group cohesion and a reduced effect of goal attraction (static floor field)Adaptive group cohesion mechanism

Modelling groups - some qualitative resultsCounterflow of two structured groups including simple groups of various size, in a 2.4 m wide corridor

Aggregate effects of groups

Counterflow of two structured groups including simple groups of various size, in a 2.4 m wide corridor; shuffled sequential update - ongoing tests with parallel update strategy

Aggregate effects of groups analysedWe can interpret the results making considering two phenomenaWide groups offer a large profile to the counter flow, so they have a higher probability of facing conflictsOnce a group has formed a line, instead, the leader has the same conflict probability of an individual, but the follower has often an advantageIn low density situations phenomenon (1) prevails, leading to a lower average combined flow for groups of pedestrians whose size is larger than 2Pairs in fact can easily form a line, turning phenomenon (1) to (2)In high density situations the probability of facing conflicts is very high also for individuals, so phenomenon (2) prevails, leading to higher average combined flow for even large groups (size 5)

Effectiveness of simple group cohesion mechanism

Counterflow of two structured groups including simple groups of various size, in a 3.6 m wide corridor(Dispersion measured in terms of area covered by the group)

Additional results in experimental scenarios: T junction

Plot of experimentally observed data[Zhang et al., 2012][Vizzari et al., 2013]


A sample simulation in the Arafat I station

Hajj case study: two vs. three groups

Hajj case study: normal vs. obstacle

Hybrid Agent Architecture

BODY

MIND

REACTIVECOGNITIVE

Per illustrare le motivazioni e lidea di architettura ibrida, utile partire che lentit modellata una persona, ed quindi naturale pensare alla parte gi esistente nel modello base come il corpo. Con il termine corpo non si vuole indicare solamente la parte di attuazione dellagente, perch infatti questo corpo implementa gi comportamenti reattivi che vanno al di l della semplice attuazioneInoltre il corpo per eseguire i suoi compiti si basa su una conoscenza di molto legata alla geometria dellambiente, proprio perch molto legato alla parte di attuazione, alla parte della fisica.

A questo punto quello che si voluto fare con questo lavoro aggiungere la parte mancante, la mente.Il termine mente risulta comodo per rendere lidea che questa parte avr a che fare con aspetti cognitivi anzich fisici, ma in realt non vuole modellare esattamente i processi mentali, bens vuole rappresentare la parte di ragionamento dellentit modellata.

Dovendo trattare aspetti pi distanti dalla rappresentazione geometrica, sar necessario dotare la mente anche di una differente conoscenza che permetta di eseguire i ragionamenti necessari.

In questo lavoro, considerata la complessit dei compiti svolti al livello tattico e strategico, ci siamo concentrati sul dotare lagente della capacit di pianificare i propri percorsi allinterno dellambiente, gettando le basi per poter sviluppare la mente in tutte le sue funzionalit.

Chiaramente, le due componenti dellarchitettura ibrida devono essere necessariamente in comunicazione, e quello che si voluto fare con questo lavoro di costruire la mente sopra al corpo, senza alterare il funzionamento precedente.35

Additional Annotation Tools

Nel modello base la fase di annotazione dellambiente permetteva di definire fondamentalmente quattro tipi di oggetti allinterno dello scenario:

Ostacoli, per modellare ad esempio i muri di una stanza;aree di partenza, dove i pedoni vengono inseriti nella simulazioneAree di destinazione, ogni pedone poteva avere una e una sola destinazione che gli veniva assegnata nel momento in cui nasceva nella simulazioneScale e Rampe, zone in cui la velocit del pedone veniva alterata per rappresentare sul piano le scale

Siccome con questo lavoro si vuole introdurre la possibilit di avere destinazioni intermedie, sono stati aggiunti anche i tool di annotazione necessari in fase di modellazione dello scenario:Regioni per strutturare lambiente da un punto di vista logico;Collegamenti di navigazione per rappresentare i passaggi tra le regioni;Target intermedi per rappresentare destinazioni intermedie di interesse

A questo punto anche target e link di navigazione sono candidati a essere destinazioni per i pedoni.36

Cognitive Map

Con le annotazioni aggiunte allo scenario, il nuovo modello in grado di calcolare in maniera automatica la mappa cognitiva.

Inoltre due delle notazioni aggiunte, potendo essere utilizzate dagli agenti come destinazioni, necessitano quindi di pathfield, cio quella parte di conoscenza dellagente che supporta il meccanismo del corpo per il raggiungimento della destinazione.

Durante la generazione di questi campi per nuovi e vecchi tipi di destinazioni, parallelamente viene costruita anche la mappa cognitiva.La sua struttura base viene generata durante la generazione dei field relativi ai link tra le regioni.

A partire da ogni marker che collega due regioni, durante la diffusione del suo campo si va alla ricerca dei marker di identificazione della regione, producendo di conseguenza una parte di mappa che viene aggiunta a quella eventualmente esistente.

Ad esempio

37

Knowledge: Body + Mind

Grazie alla generazione in parallelo , possibile quindi costruire una forte relazione tra le diverse rappresentazione dellambiente.

Nella figura mostrata proprio questa relazione tra le conoscenze, che permette di collegare la mente al corpo nonostante la distanza e differenza tra le rappresentazioni.

In questo modo, infatti, il ragionamento della mente fatto sulla mappa cognitiva pu essere facilmente tradotto in istruzioni operative che il corpo in grado di eseguire. Ad esempio dalla mappa cognitiva possibile vedere che esiste un collegamento tra sala e cucina, e che possibile istruire il corpo di effettuare questo spostamento ordinandogli di seguire il pathfield associato allarco.38

AGENT LIFE-CYCLE

Plan BuildingPlan ManagementActuation

E' tempo quindi di affrontare le modifiche che hanno interessato l'agente nel proprio modo di operare piuttosto che nella conoscenza.

In figura possibile vedere il ciclo di vita dell'agente precedente alle modifiche apportate da questo lavoro.Come si pu vedere esiste solo la parte relativa al corpo, c' una fase di percezione, che lavora sulla conoscenza del corpo, dopodich passa a selezionare il movimento reattivamente, ed infine esegue effettivamente la mossa.

Con questo lavoro sostanzialmente una fase di controllo eseguita dalla mente anticipa il precedente ciclo di vita lasciandolo praticamente inalterato.

L'idea appunto quella di lasciare il corpo lavorare come sempre, mentre la mente tiene in considerazione un piano, cambiando solamente il supporto della destinazione che il corpo vuole raggiungere.

Questo comporta 3 tipi di attivit da parte dell'agente:

Creazione di un piano, cio valutare la sequenza di destinazioni intermedie necessarie e/o obbligatorie, per raggiungere la destinazione finale di interesse.

- gestione del piano: questo il controllo che la mente esegue sul corpo, gestendo per passi il piano di destinazioni intermedie, nascondendo al corpo la presenza di un piano e fornendogli solamente l'attuale target. Il corpo infatti in grado solamente di raggiungere la destinazione che gli si ordinata, sar quindi la mente a mantenere traccia dello stato di compimento del piano e fornire di volta in volta il pathfield per il raggiungimento della destinazione intermedia corretto. Il momento in cui il piano avanza viene valutato sulla base di percezioni del corpo, che lunico in grado di percepire se si raggiunta la destinazione voluta.

- attuazione: infine c' la normale attuazione reattiva che lavora esattamente come se nessuna modifica fosse stata apportata al modello tenendo conto dei gruppi, degli ostacoli,I comportamenti reattivi39

Sample simulation considering crowd management procedures

Dynamically managing The tactical level: Introduction to the paths treeGiven an arbitrary environment, the agent should be able to plan a path toward its target, considering:The types of environment that will be crossed static elementsThe emergence of congestion or other elements influencing the path conditions dynamic elements

The choice among paths is performed according to the expected traveling time, dynamically changing.The decision tree contains the average traveling time of each minimal path to a destination, estimated by considering static elements and the average speed of the agents.

An evacuation of a large population of pedestrians1000 agents at ~ 10 p/sec

1000 agents at ~ 10 p/sec

Step 50Step 200Step 350Step 500

42

Quantitative results


Conclusions and discussionGroups are relevant and significantResults of simulations are partly validatedFundamental diagram and spatial utilisation in tune with results from the literature without groupsGroup cohesion mechanism generates results about dispersion that are in tune with Vittorio Emanuele Gallerys observation but we dont have data about groups in high density situations (and its hard to obtain such data) More observations, experiments and simulations are necessary to improve our understanding of the phenomenonCloser collaboration between researchers working on synthesis and analysis of crowds is promising and possibly beneficial for both

Future worksOf course, improvements...of the model, of our understanding about the phenomena...Of course, additional applications to real world scenarios...for providing additional success stories and collecting additional issues, limits, directions for improvementsStrengthening the connections with automated analysis/computer vision Exploiting the model for supporting smart environment, smart city systemsIn the vein of what was discussed by Ulrich Wagoum for stadium evacuation assistant, Georgios Sirakoulis work on anticipative technologies and robotic evacuation assistant...

Giuseppe Vizzari