LED wristbands for Cell-based Crowd Evacuation: an Adaptive Exit-choice Guidance System Architecture
AArticle
LED wristbands for Cell-based Crowd Evacuation: anAdaptive Exit-choice Guidance System Architecture
Miguel A. Lopez-Carmona ∗ and Alvaro Paricio Departamento de Automatica, Escuela Politecnica Superior, Universidad de Alcala, Madrid, Spain;[email protected] (M.A.L.C.); [email protected] (A.P.) * Correspondence: [email protected]; Tel.: +34-91-885-66-73Received: date; Accepted: date; Published: date
Abstract:
Cell-based crowd evacuation systems provide adaptive or static exit-choice indications thatfavor a coordinated group dynamic, improving evacuation time and safety. While a great effort has beenmade to modeling its control logic by assuming an ideal communication and positioning infrastructure,the architectural dimension and the influence of pedestrian positioning uncertainty have been largelyoverlooked. In our previous research, a Cell-based crowd evacuation system (CellEVAC) was proposedthat dynamically allocates exit gates to pedestrians in a cell-based pedestrian positioning infrastructure.This system provides optimal exit-choice indications through color-based indications and a controllogic module built upon an optimized discrete-choice model. Here, we investigate how location-awaretechnologies and wearable devices can be used for a realistic deployment of CellEVAC. We consider asimulated real evacuation scenario (Madrid Arena) and propose a system architecture for CellEVACthat includes: a controller node, a radio-controlled LED wristband subsystem, and a cell-node networkequipped with active Radio Frequency Identification (RFID) devices. These subsystems coordinate toprovide control, display and positioning capabilities. We quantitatively study the sensitivity of evacuationtime and safety to uncertainty in the positioning system. Results showed that CellEVAC was operationalwithin a limited range of positioning uncertainty. Further analyses revealed that reprogramming thecontrol logic module through a simulation-optimization process, simulating the positioning system’sexpected uncertainty level, improved the CellEVAC performance in scenarios with poor positioningsystems.Crowd evacuation; LED wristbands; behavioral optimization; exit-choice decisions;simulation-optimization modeling; cell-based evacuation
1. Introduction
Uncoordinated crowd behaviors are known as being responsible for pedestrians’ deaths and injuriesin crowd evacuations. An efficient evacuation plan is crucial to direct and coordinate evacuees in a safemanner. This coordination can be achieved by deploying guidance systems that provide information foreach user on the paths, the exit gates, or the evacuation start time [1].Many algorithms have been devised for the development of evacuation guidance systems [2]. Forexample, network flow-based approaches model evacuation planning as a minimum cost network flowproblem [3]. The main downside of network flow-based models is that evacuees must follow the pathsaccurately and fulfill an exact schedule. Various approaches have been suggested to solve this problemusing geometric graphs [4]. For example, in [5] a wireless sensor network is partitioned into triangularareas based on the average detected temperature, and safe egress paths are calculated. Following thisidea, queuing models [6] build a queuing network to estimate evacuation and congestion delays. Variousapproaches dynamically develop navigation paths by applying artificial potential fields to the exits andhazards [7–9]. This technique suffers from several problems, among which is the convergence time fornetwork stabilization, and its search efficiency in scenarios with several exit gates. a r X i v : . [ c s . OH ] A ug of 20 There is an extensive research on biologically-inspired algorithms to search for optimal routes orrecommend exits. For instance, in [10] a multiobjective evacuation route assignment model based ongenetic algorithm [11,12] is proposed. In [13] bee colony optimization is used to displace evacuees to safeareas. The downside of this work is the relatively high communication overhead. A wearable devicenamed LifeBelt is proposed in [14] that recommends exits to individuals based on the sensed environment.In [15] a shortest path algorithm computes pedestrian routes by iteratively partitioning graph edges atcritical division points. Routes are iteratively refined offline until an optimal state is achieved. Thisapproach assumes that a crowd distribution is known in advance, and does not adapt to changes duringevacuation.Since many of existing emergency response systems are built on top of Wireless Sensor Networks(WSN), routing protocols for packet transmission have been adapted to develop guidance systems inemergency scenarios. In [16] an emergency support system built on top of WSN is presented, whichis inspired by the cognitive packet network routing [17] in the IoT domain. Since communications areessential in an evacuation process, opportunistic communications have also been used in the design ofemergency support systems [18].It is well known that the performance of crowd evacuation processes during emergencies can bestrongly affected by exit-choice decision making at the individual level [19–21]. Thus, there are researchefforts in the area of real-time guidance for crowd evacuations that have focused on studying mechanismsfor providing pedestrians with optimal exit-choice information. A promising line of research in this area isthat of cell-based evacuation systems [1,22,23]. These systems rely on a cell-based pedestrian positioninginfrastructure such that pedestrians in a cell are assumed to receive the same exit gate instructions. In[1], a simulation-optimization framework integrates a genetic algorithm and a microscopic pedestriansimulation-assignment model. Evacuees are assumed to receive exit-choice indications that may includethe optimal start time of evacuation. Similarly, in [24] the idea is to use a gene expression programmingto find a heuristic rule. This rule is used to indicate people in the same sub-region to move towards thesame exit. The main drawback of these approaches is that they do not consider the dynamic environmentfeatures.Since the dynamics of the environment change over time in unpredictable ways, adaptive strategiesare recognized as more adequate solutions [21]. There exist adaptive approaches of cell-based evacuationsystems, in which the system’s control logic module updates the cells’ exit-choice indications in real-timedepending on the existing environmental conditions. In [22], they propose a heuristic rule that considersthe distance and width of exit doors as fixed input parameters and density around a given cell as adynamic parameter. The crowd evacuation planning problem is converted to finding the optimal heuristicrule that minimizes the total evacuation time. To solve this problem, the authors adopt the CartesianGenetic Programming (CGP) [25]. We developed in [23] an adaptive cell-based crowd evacuation system(CellEVAC) based on behavioral optimization that searches for the optimal evacuation plan throughmeta-heuristic optimization methodology. As in [22], we obtain adaptive evacuation plans capable ofresponding to changing environmental conditions. However, our control logic model is easier to configureand optimize, with a more straightforward logic formulation and interpretation, exhibiting a more naturalpedestrian behavior.All these approaches outlined above have mainly focused on the design of algorithms to provideoptimal exit-choice indications, assuming that there exist ideal communication and pedestrian positioninginfrastructures. However, for a real and effective implementation of this type of system, it is necessary topropose concrete hardware architectures whose deployment is technologically feasible at a reasonable cost.Also, given an architectural proposal, it will be essential to evaluate its influence on the performance of theevacuation processes, and if appropriate, propose corrective actions for its improvement. of 20
In this work, we are particularly interested in proposing an adaptive cell-based evacuation systemarchitecture using existing communication and positioning technologies, paying attention to usability,which is essential in emergency evacuations where the information of routing to exit gates should be easilyinterpretable. Another central question of this study concerns quantifying the influence of pedestrianpositioning uncertainty in evacuation time and safety. We would also like to quantify the importance ofreprogramming the control logic module under uncertainty conditions by using simulation-optimizationtechniques. Given a control logic optimized assuming an error-free positioning system, what would be thequantitative benefit of re-optimizing the control logic if we assume a level of uncertainty in pedestrianpositioning.With the purposes mentioned above, this paper proposes a system architecture for our adaptivecell-based evacuation system CellEVAC [23]. The proposed system architecture consists of threesubsystems: (i) monitoring and control logic subsystem, (ii) active RFID cell-node network, and (iii)radio-controlled LED wristband subsystem.The RFID cell-node network and radio-controlled LED wristband subsystems coordinate to provideexit gate indication display and positioning capabilities to pedestrians. The monitoring and control logicsubsystem monitors the environmental conditions and accordingly allocates exit gate colors to cell-nodesin real-time. Thus, the LED wristbands show the color corresponding to the cell in which each pedestrianis located, indicating the recommended exit gate.In [23], we assumed an error-free positioning infrastructure, where pedestrians were supposed to beequipped with a generic device (smart-phone or wearable device) with ideal location-aware and colordisplay capabilities. In this paper, the proposed positioning system’s RFID communication channels aremodeled using a log-normal propagation model. To define different uncertainty levels in pedestrianpositioning and study its influence in evacuation time and safety, we modulate the Gaussian distributionthat models the random variations in the propagation model. Finally, we apply the simulation-optimizationmethodology to obtain the control logic subsystem’s optimal configuration under different uncertaintylevels. This approach gives us information about the importance of reprogramming the control logic if weknow in advance the positioning uncertainty level.To perform simulation and simulation-optimization experiments, we have used thesimulation-optimization modeling framework that we developed in [23]. This framework integratesa microscopic pedestrian simulation based on the classical Social Force Model (SFM) [26]. Thesimulation-optimization process adopts a Tabu-Search algorithm (TS) [27], which iteratively searches forthe near-optimal evacuation plan. At the same time, the microscopic crowd simulation guides the searchby evaluating the evacuation time and safety of the solutions generated by the TS algorithm.The rest of the paper is organized as follows. Section 2 presents our proposal of system architecturefor CellEVAC, the microscopic simulation-optimization framework used to perform the experimentalevaluation, and the mechanism to model positioning uncertainty. Section 3 presents the experimentalevaluation and results. The last section provides concluding comments and possible research extensions.
2. Methods
Our evacuation scenario was
Madrid Arena , an indoor arena located in Madrid that hosts sportsevents, commercial, cultural and leisure activities. It has three floors (access, intermediate, and ground)and 30,000 m², with a capacity of 10,248 spectators. We studied the evacuation of the ground floor, whichhas a maximum capacity of 3, 400 spectators with its retractable bleachers removed. Figure 1 shows theground floor layout, with 1, 925 m and eight exit gates ( Ex1 to Ex8 ) with widths in the range 2.5 m and 6 m . of 20 Figure 1.
Madrid Arena layout (ground floor).
Pedestrian flows from intermediate floors were simulated by injecting pedestrians at exits 1, 2, 3, 4, and 6at the entry points highlighted with a blue dot.As in [23], we divided the ground floor into 42 regular hexagonal cells of 9 m and 6 m width, whosedimensions were chosen to provide a balance between control, wireless coverage, and computationalefficiency. We considered using radio-controlled LED wristbands that display colors recommending an exit gate.These LED wristbands are widely used at a range of events, from live acts at arenas to product launches,sports matches, parties, and corporate events from 1-150,000 people. The wristbands work by creatingmultiple flash patterns with RGB LEDs that use the full-color spectrum and can be programmed to createvisual effects (Figure 2). Xylobands (http://xyloband.com) or CrowdLED (https://crowdled.net) are twoexamples of companies offering these kinds of products. Usually, radio control has a range of hundreds ofmeters, and the wristbands have a battery life of approximately 20 hours. Two extended features that canbe found are the inclusion of RFID tags for registration purposes and zonal control to activate wristbandsin separate groups.Our idea was to extend the functionality of these devices, which is oriented towards creating visualeffects, using them in case of emergency to guide people to color-illuminated exit gates. The displayedcolor in the wristband indicates the evacuee the corresponding exit gate. Besides, a synergic effect of usingLED wristband lightning is that it may ease image processing for pedestrian flow estimation, which isused in our system to build the control logic. of 20
Figure 2.
As described in the introduction section, the proposed system architecture consists of threesubsystems:1. Monitoring and control logic subsystem (Controller Node), which monitors pedestrian flows usingimage processing and generates exit-choice indications in the form of color allocation to cells.2. Active RFID cell-node network whose purpose is to provide positioning information to pedestrians’LED wristbands.3. Radio-controlled LED wristband subsystem, which includes the LED wristbands with color displayand radio-frequency (RF) communication capabilities.Figure 3 shows two possible implementations of the system architecture (Types A and B) for deployingCellEVAC using existing off-the-shelf technologies. Both alternatives install a controller node with threefunctional blocks: pedestrian flow estimation based on image processing, control logic based on behavioraloptimization [23], and RF transmitter.In the controller node, the pedestrian flow estimation block performs image processing to detectLED wristbands lightning and estimate pedestrian density at each cell. In this work we assumed aflow estimation block that is based on commercially available pedestrian counting technology [28–31].Obtained pedestrian densities feed the control logic block that computes the optimal allocation of colors tocell-nodes (see Section 2.3). The RF transmitter broadcasts messages periodically containing the 42 tuples { Cell , Color } that assigns a color to each cell. This process repeats every five seconds.In the Type A architecture, each cell-node is equipped with an active RFID tag [32] that periodicallybroadcasts its ID (active RFID beacon [33]). On the other hand, the wristbands embed an RFID reader thatreads the IDs from the surrounding cells. The wristband selects the ID of the message with the highestReceived Signal Strength Indicator (RSSI) to estimate the right pedestrian location [34]. The other elementin the wristband is the RF Receiver, which periodically evaluates the broadcast messages with the tuples { Cell , Color } from the controller node. By matching the wristband location (selected cell ID) and cell-colortuples, the wristband lights up with the exit gate color assigned to the cell.In the Type B architecture, the RF Receiver in the cell-node receives the broadcast messages from thecontroller node with the assigned color. Then, the cell-node broadcasts the corresponding color message,which is read by the wristbands. As in the Type A architecture, several broadcast messages from different of 20 System ArchitectureType A
ID_CELL=N
RFID {Cell}
Pedestrianflowestimation RFID ReaderLED WristbandRF ReceiverBROADCAST to Wristbands
Map {Cell, Color}
Image Processing Controllogic
X (input): d Y (input): n Z (output): scoreX grids: 15 Y grids: 15 EvaluatePlot points:Ref. Input: 101 Help CloseSpecify a FIS file name.
Controller Node
ColorWristbandled lights
System ArchitectureType B
ID_CELL=N {Color}
Pedestrianflowestimation RF ReceiverLED WristbandRF ReceiverImage Processing Controllogic
X (input): d Y (input): n Z (output): scoreX grids: 15 Y grids: 15 EvaluatePlot points:Ref. Input: 101 Help CloseSpecify a FIS file name.
Controller Node
ColorBROADCAST to CellNodes
Map {Cell, Color}
RF TransmitterWristbandled lights
Figure 3.
CellEVAC System Architecture: Types A and B. of 20 cell-nodes can be received within a window time. So, the same signal strength selection mechanism isused by the wristbands to select the right color.The most critical part of this architecture is in the positioning functionality. Both RFID and RFcommunication channels between the cell-node network and wristbands have to cope with a complexsignal propagation environment. However, the system does not need to obtain exact position coordinatesbut select the right cell in which the pedestrians are located. It means that a significant lower locationresolution is needed and that there is no need to implement triangulation mechanisms based on RSSI [35].Another problem to solve is co-channel interference, which may be managed using existing radio resourcemanagement schemes [36]. Besides, the RF transmission channel in the controller node is a one-to-manycommunication channel that has been used to control commercially available LED wristbands in largeevents for more than a decade, and do not pose a particular challenge.
The control logic of CellEVAC is based on a behavioral optimization approach proposed in ourprevious work in [23]. Here we recall the main concepts that build its operation.Pedestrians’ exit-choice decision modeling based on discrete choice theory [37] inspired the controllogic developed for CellEVAC. Thus, we modeled exit gate color allocation to cell-nodes using the simplestand most popular practical discrete choice model, the Multinomial Logit Model (MLM) [37,38]. In theMLM control logic, the controller node allocates exit gates (colors) to cells using a probabilistic model, inwhich the allocation probabilities of exit gate j to cell-node c are given by: P cj = exp ( V cj ) ∑ E i ∈ E ( c ) exp ( V ci ) (1)where E = { E i = } is the set of exit gates, and V cj is the systematic utility function for cell c and exit gate j , which is given by: V cj = β D × DISTANCE cj max ( DISTANCE ) + β W × W IDTH j max ( W IDTH ) (2) + β G × GROUP cj − GROUP min
GROUP cj + β E × EXCON j criticalDensity j + β C ( t ) × NOCH ANGI NG cj The first attribute is the distance from cell-node c to exit gate j , while the second attribute representsthe width of each exit gate. Both attributes are normalized in the range of 0-1.The third attribute is the GROUP ratio, which estimates the congestion along a path from cell c toan exit gate j , relative to the least congested path. A group ratio value of 0 indicates that the chosen pathis the least congested. When the value of the group ratio tends towards 1, it means that the emptiestpath’s imbalance becomes large. The parameter β G is expected to be positive if it favors pedestrians tofollow other pedestrians and is negative otherwise. Note that with this normalization, we assume that theattribute’s relevance is kept constant throughout the evacuation process.The fourth attribute EXCON accommodates the congestion at exit gates. For a given density value,congestion is higher if the pedestrian flow is low. We reflect this effect through critical density valuesobtained from the fundamental diagrams of each exit gate (see [23]). This criticalDensity j value reflects the of 20 density value at which the exit gate’s maximum capacity is reached. Therefore, the EXCON j valuerepresenting density at exit gate j is normalized by the corresponding criticalDensity j value. Thisnormalization converts EXCON into a unitless attribute around 1. When the value of
EXCON is above 1,it means that exit is highly congested. A value close to 0 would indicate that the exit gate is almost empty.In contrast to the normalization procedure used for the
GROUP attribute, the distribution of
EXCON values exhibits a decreasing evolution as the number of pedestrians in the evacuation scenario decreases. Itseems reasonable to assume that the relevance of congestion at exits as a discriminant factor for exit-choicedecreases when the overall number of pedestrians is low, and so
EXCON is close to 0 at all exits.The fifth attribute is the
NOCH ANGI NG value associated with cell-node c and exit j , which captureshow the controller is likely to revise the previous exit gate allocation (this attribute was named PERSON AL in [23]) . We treat this attribute as a binary categorical 0-1 value that equals 1 if the current exit gate of cell c is j , and is 0 otherwise ( NOCH ANGI NG = ∀ k (cid:54) = j ). Therefore, in a general context, the parameter β C ( t ) is expected to be positive if the controller tends to maintain the previous exit gate allocations, andis negative otherwise. However, we aimed to modulate the tendency to maintain previous decisions,and so, β C ( t ) is always positive. As was noted above, we assumed that exit-choice changing evolves asevacuation progresses, and therefore the parameter that modulates NOCH ANGI NG is time-dependent.By observing the pattern of behavior under various simulation settings, it was found reasonable that thetendency to maintain decisions increased linearly depending on the current number of pedestrians: β C ( t ) = β C × (cid:18) − numO f Peds ( t ) numO f Peds ( t = ) (cid:19) (3)According to Equation 3, the parameter β C ( t → ) tends to 0 at the beginning of the evacuation, andso, the tendency to revise decisions is higher. As the number of pedestrians decreases, the parameter β C ( t ) tends to β C , and the tendency is to maintain previous decisions proportionally to the β C value.In the simulation setting used in this work, we used an update cycle of 5 seconds. We kept thisfrequency constant and controlled the frequency of the changes at optimal levels using the parameter β C . Active RFID systems are defined by three parts, a reader (wristband), antennas, and a tag (cell-node),with their power source. In active RFID applications, RSSI can be used for determining the location of atag, such that each tag’s RSSI value is proportional to the distance. In our system, the cell-node embeds anactive beacon tag that sends out its ID every 3 - 5 seconds. Thus, each tag’s RSSI value is proportional tothe distance between the wristband and cell-node. However, the RSSI value in active RFID applicationscan be affected by multipath and signal collision [35].In free-space, the relationship of the power transmitted from cell-node to wristbands, assuming theantennas are isotropic and have no directivity, is expressed by the free-space path loss equation derivedfrom the Friis transmission equation: PL ( dB ) =
20 log ( d ) +
20 log ( f ) − PL is the free-space path loss in dB , f is the signal frequency in MHz, and d is the distance in metersfrom the cell-node to the wristband. For converting RSSI values into a distance measurement in indoorenvironments with random shadowing effects, one of the most common approaches taken is to use thelog-normal propagation model [39,40]: P RX dBm = RSSI = P TX dBm − PL − η log dd + X g (5) of 20 where P TX dBm is the transmitted power in dBm , P RX dBm is the received power, PL is the path lossfor a reference distance d calculated using the free-space path loss equation (Equation 4) or by fieldmeasurements, d ≥ d is an arbitrary distance, η is the path loss exponent, and X g is a gaussian randomvariable with zero mean and variance σ that models the random variation of the RSSI value. The pathloss exponent η in indoor environments such as stadiums can reach values in the range of 4 to 7.User preference or environmental considerations usually prescribe which parameter configurationto use for most applications. In our simulation scenario, we used a frequency of 2.45 GHz , transmissionpower of 10 W , path loss exponent η =
5, and reference distance d = m . Thus, RSSI can be expressed as RSSI = −
60 log ( d ) + X g , d ≥ m (6)Modifying the variance σ g of the gaussian distribution X g we may modulate positioning uncertainty.The procedure to calculate each pedestrian’s location in evacuation simulations is a two-step processthat repeats every five seconds:1. Given the set of cell-nodes { c i = } obtain the set of distances { d i = } from pedestrian to eachcell-node c i , and calculate the corresponding set { RSSI i } using Equation 6.2. If there exists a distance value d i in set { d i = } such that d i < m , the pedestrian location is c i .Otherwise, the pedestrian location corresponds to the cell c i with the maximum RSSI i value. Much of the related work on crowd evacuations rely on detailed simulations. We opted for amulti-agent microscopic simulation framework based on a Social Force Model (SFM) [26] due to itsflexibility and ease of integration of complex interaction and behavior models. Our simulation frameworkintegrates the potential of SFM to mimic physical interactions among evacuees, and of multi-agent systemsto simulate complex behaviors and interactions [41].In this work, the simulation-optimization software framework we developed in [23] has been extendedwith the positioning uncertainty model. The software framework embeds agent-based simulation anddiscrete event simulation, integrating pedestrian behavior modeling, SFM for pedestrian motion, controllogic of exit gate indications, positioning, and optimization features.We used the commercially available programming, modeling and simulation software packagesAnyLogic and Matlab . The kernel of the simulation-optimization software framework is AnyLogic,which integrates three different modeling methods: discrete event simulation, agent-based simulation,and system dynamics, built on top of a Java-based software development framework. The evacuationscenario layout, pedestrian motion, and evacuation measurements run in AnyLogic, while exit-choicedecisions and control logic of exit gate indications are implemented in Matlab. AnyLogic and Matlab areinterconnected in a master-slave configuration through the interface with external Java libraries providedby AnyLogic and the Matlab Java API engine (see details below).The CellEVAC simulation model with MLM control logic is shown in Figure 4. The evacuationscenario layout, visualization features, and all the functionality regarding the SFM based pedestrianmotion were implemented within AnyLogic.During a simulation, the first step is to send from AnyLogic to Matlab the set of parameters thatconfigure the CellEVAC MLM and Pedestrians’ positioning modules, including the set of cell-node center Tabu-searchOptimizationProcess
MLM candidates
Simulation results foreach MLM candidate M A T L AB J a v a AP I X (input): d Y (input): n Z (output): scoreX grids: 15 Y grids: 15 EvaluatePlot points:Ref. Input: 101 Help CloseSpecify a FIS file name.
Mappingof coloursto cells
SET {Colour/Exit(Cell01),...,Colour/Exit(Cell42) }
SET {GroupSize(Cell01,Ex1),..., GroupSize(Cell01,Ex8),..., GroupSize(Cell42,Ex1),...,GroupSize(Cell42,Ex8) }SET { Density(Ex1),..., Density(Ex8) }
Attributes
Matlab
CellEVAC MLM
EXIT 1 EXIT 2 … EXIT 8CELL 1 CELL 2 … … … … … … CELL 42
MATLAB Engine M A T L AB J a v a AP I Periodic measurements
Pedestrians’ positioning
Pedestrians’ exit-choice
SET {Exit(P ),...,Exit(P ),...,Exit(P n ) } PEDESTRIANS’ POSITIONS
Figure 4.
Simulation-optimization software framework of CellEVAC with control logic based onMultinomial Logit Model (MLM). coordinates and exit gates, and the uncertainty level. Next, the pedestrian positioning and densities atexit gates and cells are periodically measured and then transformed into the set of attributes: pedestrianpositions, density at each exit gate, and group of pedestrians along the path to each exit. The group sizeof each pair cell-exit gate is calculated by adding the pedestrians in the cells that are closer to each exit.All these attributes feed the CellEVAC MLM module in Matlab that implements the decision logic tomap colors (exit gates) to cells. This mapping is sent back to AnyLogic for visualization purposes, andto the Pedestrians’ positioning module within Matlab to allocate exit gates (colors) to pedestrians (LEDwristbands). The Pedestrians’ positioning module implements the function that locates each pedestrianin a given cell-node using the positioning uncertainty model. The output of the Pedestrians’ positioningmodule is the set of pairs pedestrian-exit gate, which is sent to AnyLogic for simulating pedestrian motion.To search for optimal configurations of the MLM model, we used a simulation-optimization processthat adopts a Tabu-Search algorithm (TS) [27], which iteratively searches the solution space. At the sametime, the microscopic crowd simulation guides the search by evaluating the evacuation time and safetyof the solutions generated by the TS algorithm. The optimization process is built on top of the OptQuest optimization engine provided by AnyLogic. Figure 4 shows the optimization module on a greenbackground. The parameters of the CellEVAC MLM model are the “MLM candidates” generated by the TSalgorithm. Thus, each candidate is defined by a tentative set of parameters β sent to the MATLAB Engineat each iteration of the optimization process. The simulations results are sent back to the optimizationmodule for its evaluation and thus guide the optimization process.
3. Simulation-optimization Experiments and Results
The performance measurements in all the experiments were the total evacuation time , average safety,variance of safety , and the average number of pedestrians’ exit-choice decision changes . The average and varianceof safety are based on the safety values computed at the differents exit gates. Average safety characterizesthe overall safety value, while the variance of safety is used to estimate the imbalance of safety betweenthe exit gates. The procedure to calculate exits’ safety values is first to obtain their Fundamental Diagrams(FD) derived through microscopic simulation. The FD represents the relation between pedestrians’ flowand density. Given the FDs, a procedure is defined to obtain three density thresholds. These densitythresholds and the measurements of density during an evacuation process are used to calculate the safetyvalues. For a detailed description of how we modeled pedestrian flows and safety, see [23]. In this paper,we used the same thresholds and parameters defined in [23] to measure the safety values.We conducted two types of experiments: (i) sensitivity analysis to positioning uncertainty, and (ii)simulation-optimization. In all the simulation setups, the evacuee population consisted of 3400 pedestrianson the ground floor, who had a preferred evacuation speed obtained from a uniform distribution between1.24 and 1.48 m / s . To speed up the simulation-optimization experiments, we imposed a deadline of25minutes to each evacuation simulation iteration, after which the simulation iteration was aborted.Two different evacuation scenarios were considered depending on the experiment: evacuationswithout external flows (NEF) in which no pedestrians were coming from the upper floors, and evacuationswith external flows (EF) (i.e., with pedestrians coming from the upper floors) to simulate more complexpedestrian flow interactions. In EF scenarios, three exit gates were chosen at random at each simulationiteration. Two of these exit gates received an incoming pedestrian flow rate of 120 peds / m , while the thirdexit gate was blocked.In the sensitivity analysis experiments, each experiment ran the evacuation simulation model multipletimes varying the positioning uncertainty level (variance of the Gaussian distribution X g ), showing howthe simulation output (i.e., the performance measurements) depended on it. Due to the stochastic natureof the evacuation processes, we used a replication algorithm to obtain representative results for a givenparameter setting and a specific simulation output. This algorithm defines a minimum and a maximumnumber of experimental runs per parameter setting (replications of a simulation), a confidence level forthe sample mean of replications (simulation output average), and an error percent. The minimum numberguarantees the minimum number of replications, while the confidence level and error percent determine ifmore replications are needed. Simulations for a given parameter configuration stops when the maximumnumber of replications has been run or when the confidence level is within the given percentage of themean of the replications to date. In our experimental setup, evacuation time was used as an outputparameter to control the number of replications between 10 and 50. The confidence level was fixed to 95%,and the error percent to 0.5.In the simulation-optimization experiments, we used the Tabu-search optimization algorithm [27].The goal was to find the optimal combination of parameters of the MLM model that resulted in the bestpossible solution. We considered two different optimization scenarios characterized by the fitness function(objective function) used and the existence of external pedestrian flows.• NEF : Optimize Time and Safety ( min ( evacTime − S f ) ) without External Flows. The goal is tooptimize evacuation time and average safety, and the evacuation scenario does not include externalpedestrian flows.• EF : Optimize Time and Safety ( min ( evacTime − S f ) ) with External Flows. The goal is to optimizeevacuation time and average safety, and the evacuation scenario includes external pedestrian flows.As in the sensitivity analysis experiments, the optimization algorithm applies a replication algorithm.However, while in the sensitivity analysis, the number of replicas was limited by the evacuation time Table 1.
Optimal parameter configuration for the CellEVAC MLM decision logic model, and parameterconfiguration of the MLM pedestrians’ standard behavior model. β D β G β E β W β C Optimal CellEVAC for 0dB -17.723 -2.181 -1.671 1.064 2.594Standard (without CellEVAC) -28 0.6 -0.5 0.6 0 value, in simulation-optimizations, the stop condition was controlled by the fitness function (objectivefunction).
In the sensitivity analysis of evacuation performance to positioning uncertainty, the standard deviation σ g in X g (Equation 6) was varied from 0dB to 40dB at discrete steps in two different evacuation scenarios,with and without external pedestrian flows. To evaluate up to which uncertainty level is beneficialCellEVAC in comparison with not using a guidance system, we also included the case in which pedestriansdid not use the CellEVAC system (coded as σ g = N in the result box-plots). The optimal parameterconfiguration of the MNL model found in [23] was used to implement the CellEVAC decision logic (Table1).. For the experiments in which pedestrians did not use CellEVAC, the decision logic was implemented ata pedestrian level using the configuration of parameters of the MLM model defined in [23] that simulatesstandard pedestrian behavior.For illustration purposes, Figure 5 shows still images 25 seconds after the start of the evacuationfor different standard deviation values σ g , from 0dB to 40dB. As expected, the snapshots exhibited aprogressive level of error in cell detection, becoming more evident from 15dB.In evacuation scenarios without external flows (Figure 6), results revealed that evacuation timeincreased linearly for σ g above 5dB. Regarding safety, increasing values of σ g had a significant negativeimpact on average safety, though for σ g above 10dB average safety stabilized around −
15. Besides, theimpact on safety variance was not so significant as in average safety. As expected, performance worsenedfor increasing σ g , though for values above 20dB safety variance tended to improve and stabilize. At thecost of an increasing evacuation time, we observed how uncoordinated pedestrians’ movement whenpositioning uncertainty was high, made spatial-density at exit gates decrease, and so safety measurementsstabilize or improve. The number of exit-choice decision changes increased exponentially with σ g , due tothe logarithmic scale (dB) used to define the values of σ g .When compared to evacuations without external flows in which CellEVAC did not operate, and notconsidering the safety variance, we observed that using CellEVAC was beneficial strictly for values of σ g below 5dB. Higher values of σ g could be valid at the cost of an increase in evacuation time. However, notethat not using CellEVAC comes at the cost of a significantly higher safety variance.In evacuation scenarios with external flows, the sensitivity analysis results revealed the same trend aswithout external pedestrian flows (Figure 7). When compared to evacuations that did not use CellEVAC,and not considering the safety variance, the benefit of CellEVAC expanded to σ g below 10dB. However,note that safety variance is exceptionally high when not using CellEVAC, which means that pedestrianflows at different exit gates is highly unbalanced.Overall, the results of the sensitivity analyses for scenarios with and without external flows suggest aclear benefit of using CellEVAC if the positioning system exhibits RSSI random variations below 10dB. Figure 8 shows the progress of the Tabu-search simulation-optimization of the MLM models’parameter configurations for values of σ g equal to 5, 15 and 20dB. The objective was to optimize Figure 5.
Still images 25 seconds after the start of the evacuation from single run simulation experimentsfor different standard deviation values σ g . The cells are shaded with the exit-gate color allocated by thecontroller node. Colored dots represent pedestrians with the colors shown by their LED wristbands. N0 5 10 15 20 25 30 35 40 g (dB) E v a c ua t i on t i m e ( m ) N0 5 10 15 20 25 30 35 40 g (dB) -20-19-18-17-16-15-14-13-12-11-10 A v e r age s a f e t y N0 5 10 15 20 25 30 35 40 g (dB) S a f e t y v a r i an c e N0 5 10 15 20 25 30 35 40 g (dB) E x i t - c ho i c e de c i s i on c hange s Figure 6.
Sensitivity analysis of the positioning uncertainty in evacuation scenarios without externalpedestrian flows. The box-plots on the first row show the sensitivity of evacuation time and averagesafety to standard deviation values σ g in the range 0dB to 40dB. The second-row plots show the sensitivityof safety variance and the number of decision changes to the standard deviation values σ g . In the fourbox-plots, σ g = N represents an evacuation scenario in which pedestrians do not use CellEVAC. N0 5 10 15 20 25 30 35 40 g (dB) E v a c ua t i on t i m e ( m ) N0 5 10 15 20 25 30 35 40 g (dB) -26-24-22-20-18-16-14-12 A v e r age s a f e t y N0 5 10 15 20 25 30 35 40 g (dB) S a f e t y v a r i an c e N0 5 10 15 20 25 30 35 40 g (dB) E x i t - c ho i c e de c i s i on c hange s Figure 7.
Sensitivity analysis of the positioning uncertainty in evacuation scenarios with external pedestrianflows. The box-plots on the first row show the sensitivity of evacuation time and average safety to standarddeviation values σ g in the range 0dB to 40dB. The second-row plots show the sensitivity of safety varianceand the number of decision changes to the standard deviation values σ g . In the four box-plots, σ g = N represents an evacuation scenario in which pedestrians do not use CellEVAC. Iteration E v a c ua t i on ( m ) - S a f e t y Iteration E v a c ua t i on ( m ) - S a f e t y Iteration E v a c ua t i on ( m ) - S a f e t y Figure 8.
Progress of the Tabu-search simulation-optimization of the MLM models that implement theCellEVAC guidance system for σ g equal to 5, 15 and 20dB. Solutions below the current best solution (redline) correspond to non-viable solutions. Table 2.
Optimal parameter configurations of the MLM model for different values of σ g . σ g β D β G β E β W β C evacuation time and average safety in scenarios with external flows. It was assumed that the entirepopulation of evacuees followed the indications of the CellEVAC system. Also, we imposed an arbitrarysimulation stop-limit of 25 minutes to evacuation time, and a restriction to the viability of the solutionswas incorporated to remove solutions in which there were pedestrians pending evacuation.The optimal parameter configurations found are shown in Table 2. We did not observe significantdifferences between the different parameter configurations, except for the β D and β G parameters. Thedistance parameter gained more influence in scenarios with high positioning uncertainty. Remarkably, for σ g = β G had a positive sign. Our interpretation is that a higher influence ofdistance and a positive value of β G contributes to more uniformity in the exit gate indications and lessscattering in color allocation to cells. As a consequence of this, the probability of location error decreases.The optimal configurations found were tested in evacuations with different positioning uncertaintylevels, from 0dB to 40dB in steps of 5dB. The results have been summarized in Figures 9 and 10 underevacuation scenarios without and with external flows, respectively.In evacuation scenarios without external flows (Figure 9), average evacuation time, and exit-choicedecision change performance measurements did not show significant differences between the differentconfigurations and evacuation scenarios. Interestingly, we found a positive trend in the results in terms ofaverage safety and safety variance for evacuation scenarios for 20dB and above when using the optimalconfiguration found for 20dB. In evacuation scenarios with external flows (Figure 10), the results weresimilar except for safety variance, for which we did not observe any improvement.The results presented in section 3.1 show that CellEVAC is useful only if the positioning systemexhibits RSSI random variations below 10dB. Besides, the performance analysis results of the optimalconfigurations do not exhibit any improvement below 20dB. Consequently, we can conclude that there isno evidence that optimizing the MLM model under the assumption of an expected random variance ofRSSI contributes to an improvement in the performance of the CellEVAC system. Figure 9.
Median box-plots of the performance measurements of the optimal configurations of CellEVACfor σ g =
0, 5, 10, 20dB, in evacuation scenarios without external pedestrian flows and with positioninguncertainty levels from 0dB to 40dB in steps of 5dB. Bottom horizontal axes categorize the optimalconfiguration of parameters used (0, 5, 10, or 20dB) to configure CellEVAC. Upper horizontal axes categorizethe σ g value that models the positioning system in the evacuation scenario. For instance, a value of 25dB inthe axis “ σ g (dB) in evacuation scenario” and 15dB in “Optimized in σ g (dB)” expresses that CellEVAC hasbeen configured to use the optimal configuration found with 15dB, and that it is tested in an evacuationscenario with σ g = Figure 10.
Median box-plots of the performance measurements of the optimal configurations of CellEVACfor σ g =
0, 5, 10, 20dB, in evacuation scenarios with external pedestrian flows and with positioninguncertainty levels from 0dB to 40dB in steps of 5dB. Bottom horizontal axes categorize the optimalconfiguration of parameters used (0, 5, 10, or 20dB) to configure CellEVAC. Upper horizontal axes categorizethe σ g value that models the positioning system in the evacuation scenario.
4. Conclusion
Our use of an MLM model to implement the control logic of Cell-based crowd evacuation systemshas proven to be very efficient (see [23]). However, as in other existing works on cell-based crowdevacuation systems [1,15,24], little attention has been paid to propose a system architecture based onexisting technologies and assuming real communication and positioning infrastructures. In our opinion,these considerations are crucial to boost the real implementation of these systems.In this paper, we have proposed a specific system architecture built upon radio-controlled LEDwristbands that connect with a cell-node network and a controller node, through radio-frequencycommunication channels. The use of LED wristbands has numerous advantages, among which wehighlight its low cost, being a technology widely used to create visual effects at large events, and beingan intuitive and straightforward interface that can make it exceptionally efficient in emergencies. Thistype of indication system greatly simplifies the interpretation of exit gate indications, which is particularlyimportant in stressful situations found typically in evacuation scenarios. Indirectly, it can also improve theaccuracy of the detection of pedestrian flows in the controller node.Another of our aims was to quantitatively study the sensitivity of evacuation time and safety touncertainty in the positioning system. With this objective, we have modeled the communication channelbetween the LED wristbands and the cell-nodes using a log-normal propagation model. To modeldifferent levels of uncertainty in positioning, we have modulated the random variations of RSSI fromthe propagation model. In the sensitivity analysis of performance parameters to different values of RSSIvariance, CellEVAC is shown to be operational strictly up to values of 10dB. The system generates toomany color changes in the wristbands and a significant increase in evacuation times for higher values.Our last goal was to evaluate if it was possible to improve the CellEVAC performance obtaining newoptimal MLM parameter configurations in which different levels of RSSI standard deviation were assumed.The results obtained have revealed that improvements found are relevant only for evacuation scenarioswith levels of positioning uncertainty greater than 20dB, in which CellEVAC is not operational. Thus, tooptimize the MLM model used in the CellEVAC control logic, it is valid to assume that the positioningsystem is error-free. However, the system cannot be applied in a real environment if the standard deviationof the RSSI values is greater than 10dB.Several extensions are being considered for this research, mainly focused on investigating howto expand the allowed range of RSSI variation without the need to modify the existing positioninginfrastructure. Another research extension is related to developing a prototype of the CellEVAC systemarchitecture proposed in this paper.
Author Contributions:
Conceptualization, M.A.L.C; methodology, M.A.L.C and A.P.; software, M.A.L.C.; validation,A.P. and M.A.L.C; investigation, M.A.L.C and A.P.; resources, M.A.L.C; data curation, M.A.L.C and A.P.;writing–original draft preparation, M.A.L.C.; writing–review and editing, M.A.L.C and A.P.; visualization, A.P.and M.A.L.C; supervision, M.A.L.C; project administration, M.A.L.C; funding acquisition, M.A.L.C.
Funding:
This work was supported in part by the Spanish Ministry of Economy, Industry, and Competitiveness underGrant TIN2016-80622-P (AEI/FEDER, UE).
Conflicts of Interest:
The authors declare no conflict of interest.1. Abdelghany, A.; Abdelghany, K.; Mahmassani, H.; Alhalabi, W. Modeling Framework for Optimal Evacuationof Large-Scale Crowded Pedestrian Facilities.
European Journal of Operational Research , , 1105–1118.doi:10.1016/j.ejor.2014.02.054.2. Bi, H.; Gelenbe, E. A Survey of Algorithms and Systems for Evacuating People in Confined Spaces. Electronics , , 711. doi:10.3390/electronics8060711.
3. Ford, L.R.; Fulkerson, D.R.
Flows in Networks ; Princeton University Press, 1962.4. Li, X.; Calinescu, G.; Wan, P.J.; Wang, Y. Localized Delaunay Triangulation with Application in Ad Hoc WirelessNetworks.
IEEE Trans. Parallel Distrib. Syst. , , 1035–1047. doi:10.1109/tpds.2003.1239871.5. Chen, P.Y.; Chen, W.T.; Shen, Y.T. A Distributed Area-Based Guiding Navigation Protocol for Wireless SensorNetworks. 2008 14th IEEE International Conference on Parallel and Distributed Systems, 2008, pp. 647–654.doi:10.1109/ICPADS.2008.80.6. Newell, C. Applications of Queueing Theory ; Vol. 4, Springer Science & Business Media, 2013.7. Koditschek, D. Robot Planning and Control Via Potential Functions.
The Robotics Review , pp. 349–367.8. Hill, J.; Szewczyk, R.; Woo, A.; Hollar, S.; Culler, D.; Pister, K. System Architecture Directions forNetworked Sensors. Proceedings of the Ninth International Conference on Architectural Support forProgramming Languages and Operating Systems; ACM: New York, NY, USA, 2000; ASPLOS IX, pp. 93–104.doi:10.1145/378993.379006.9. Li, Q.; De Rosa, M.; Rus, D. Distributed Algorithms for Guiding Navigation Across a Sensor Network.Proceedings of the 9th Annual International Conference on Mobile Computing and Networking; ACM: NewYork, NY, USA, 2003; MobiCom ’03, pp. 313–325. doi:10.1145/938985.939017.10. Li, Q.; Fang, Z.; Li, Q.; Zong, X. Multiobjective Evacuation Route Assignment Model Basedon Genetic Algorithm. 2010 18th International Conference on Geoinformatics, 2010, pp. 1–5.doi:10.1109/GEOINFORMATICS.2010.5567485.11. Holland, J.H.
Adaptation in Natural and Artificial Systems: An Introductory Analysis with Applications to Biology,Control and Artificial Intelligence ; MIT Press: Cambridge, MA, USA, 1992.12. Gelenbe, E.; Liu, P.; LainLaine, J. Genetic Algorithms for Route Discovery.
IEEE Transactions on Systems, Man,and Cybernetics, Part B (Cybernetics) , , 1247–1254. doi:10.1109/TSMCB.2006.873213.13. Yadegari, M. A Biologically-Inspired Optimization Algorithm For Urban Evacuation Planning In DisasterManagement. 2010.14. Ferscha, A.; Zia, K. LifeBelt: Crowd Evacuation Based on Vibro-Tactile Guidance. IEEE Pervasive Computing , , 33–42. doi:10.1109/MPRV.2010.83.15. Wong, S.K.; Wang, Y.S.; Tang, P.K.; Tsai, T.Y. Optimized Evacuation Route Based on Crowd Simulation. Computational Visual Media , , 243–261. doi:10.1007/s41095-017-0081-9.16. Filippoupolitis, A. Emergency Simulation and Decision Support Algorithms. PhD thesis, 2010.17. Nowak, M.; Nowak, S.; Doma ´nska, J.; Czachórski, T. Cognitive Packet Networks for the Secure Internet ofThings. 2019 Global IoT Summit (GIoTS), 2019, pp. 1–4. doi:10.1109/GIOTS.2019.8766380.18. Gorbil, G.; Gelenbe, E. Opportunistic Communications for Emergency Support Systems. Procedia ComputerScience , , 39–47. doi:10.1016/j.procs.2011.07.008.19. Kinateder, M.; Comunale, B.; Warren, W.H. Exit Choice in an Emergency Evacuation Scenario Is Influenced byExit Familiarity and Neighbor Behavior. Safety Science , , 170–175. doi:10.1016/j.ssci.2018.03.015.20. Chen, L.; Tang, T.Q.; Huang, H.J.; Song, Z. Elementary Students’ Evacuation Route Choice in a Classroom:A Questionnaire-Based Method. Physica A: Statistical Mechanics and its Applications , , 1066–1074.doi:10.1016/j.physa.2017.11.036.21. Haghani, M.; Sarvi, M. Simulating Dynamics of Adaptive Exit-Choice Changing in Crowd Evacuations: ModelImplementation and Behavioural Interpretations. Transportation Research Part C: Emerging Technologies , , 56–82. doi:10.1016/j.trc.2019.04.009.22. Zhong, J.; Cai, W.; Luo, L. Crowd Evacuation Planning Using Cartesian Genetic Programming and Agent-BasedCrowd Modeling. 2016, Vol. 2016-February, Proceedings - Winter Simulation Conference , pp. 127–138.doi:10.1109/WSC.2015.7408158.23. Lopez-Carmona, M.A. CellEVAC: An Adaptive Guidance System for Crowd Evacuation through BehavioralOptimization. arXiv:2007.05963 [cs, eess] , [arXiv:cs, eess/2007.05963].24. Zhong, J.; Luo, L.; Cai, W.; Lees, M. Automatic Rule Identification for Agent-Based Crowd Models throughGene Expression Programming. 2014, Vol. 2, , pp. 1125–1132.
25. Miller, J. Cartesian Genetic Programming.
Cartesian Genetic Programming .26. Helbing, D.; Molnár, P. Social Force Model for Pedestrian Dynamics.
Physical Review E , , 4282–4286.doi:10.1103/PhysRevE.51.4282.27. Fred Glover. Tabu Search ; Kluwer Academic Publishers, 1997.28. Lesani, A.; Nateghinia, E.; Miranda-Moreno, L.F. Development and Evaluation of a Real-Time PedestrianCounting System for High-Volume Conditions Based on 2D LiDAR.
Transportation Research Part C: EmergingTechnologies , , 20–35. doi:10.1016/j.trc.2020.01.018.29. Kurkcu, A.; Ozbay, K. Estimating Pedestrian Densities, Wait Times, and Flows with Wi-Fi and BluetoothSensors. Transportation Research Record , , 72–82. doi:10.3141/2644-09.30. Akhter, F.; Khadivizand, S.; Siddiquei, H.R.; Alahi, M.E.E.; Mukhopadhyay, S. IoT Enabled IntelligentSensor Node for Smart City: Pedestrian Counting and Ambient Monitoring. Sensors , , 3374.doi:10.3390/s19153374.31. Ilyas, N.; Shahzad, A.; Kim, K. Convolutional-Neural Network-Based Image Crowd Counting: Review,Categorization, Analysis, and Performance Evaluation. Sensors , , 43. doi:10.3390/s20010043.32. Chen, Y.L.; Liu, D.; Wang, S.; Li, Y.F.; Zhang, X.S. Self-Powered Smart Active RFID Tag Integrated with WearableHybrid Nanogenerator. Nano Energy , , 103911. doi:10.1016/j.nanoen.2019.103911.33. Correa, A.; Barcelo, M.; Morell, A.; Vicario, J.L. A Review of Pedestrian Indoor Positioning Systems for MassMarket Applications. Sensors , , 1927. doi:10.3390/s17081927.34. Alvarez Lopez, Y.; de Cos Gomez, M.E.; Las-Heras Andres, F. A Received Signal Strength RFID-Based IndoorLocation System. Sensors and Actuators A: Physical , , 118–133. doi:10.1016/j.sna.2017.01.007.35. Brchan, J.L.; Zhao, L.; Wu, J.; Williams, R.E.; Pérez, L.C. A Real-Time RFID Localization ExperimentUsing Propagation Models. 2012 IEEE International Conference on RFID (RFID), 2012, pp. 141–148.doi:10.1109/RFID.2012.6193042.36. Assarian, A.; Khademzadeh, A.; Hosseinzadeh, M.; Setayeshi, S. An Efficient Resource Allocation Scheme in aDense RFID Network Based on Cellular Learning Automata. International Journal of Communication Systems , , e3835. doi:10.1002/dac.3835.37. Duives, D.C.; Mahmassani, H.S. Exit Choice Decisions during Pedestrian Evacuations of Buildings. Transportation Research Record , , 84–94. doi:10.3141/2316-10.38. Ortúzar, J.; Willumsen, L. Modelling Transport , fourth ed.; John Wiley and Sons: New York, 2011.39. Rappaport, T.S. Wireless Communications – Principles and Practice, Second Edition. (The Book End).