CellEVAC: Adaptive Crowd Evacuation System

• CellEVAC is an adaptive evacuation guidance system that partitions venues into cells to dynamically recommend exits based on real-time congestion and flow.
• It uses a deterministic, multinomial logit model to balance exit usage by integrating pedestrian behavior, local density, and exit attributes.
• Deployment integrates RFID cell nodes and LED wristbands, with simulation-optimization showing enhanced safety and efficacy even with moderate compliance.

CellEVAC, expanded by its authors as Cell-based Crowd Evacuation, is an adaptive evacuation guidance system for emergency egress in multi-exit venues. It partitions a facility into spatial cells and dynamically assigns each cell a color corresponding to a recommended exit, so that evacuees in the same cell receive a common, location-dependent exit-choice indication rather than a static nearest-exit instruction. The system was introduced as a behaviorally grounded alternative to both static signage and opaque heuristic controllers, with the explicit aim of improving evacuation time and safety under evolving congestion, blocked exits, and interfering pedestrian flows (Lopez-Carmona et al., 2020). A subsequent architecture paper extended the concept toward deployment using a controller node, active RFID-equipped cell nodes, and radio-controlled LED wristbands, and examined the effect of localization uncertainty on system performance (Lopez-Carmona et al., 2020).

1. Concept and problem setting

CellEVAC addresses a familiar failure mode in large venues with multiple exits: if evacuees simply head toward the nearest exit or follow the crowd, some exits become overloaded while others are underused, producing bottlenecks, unstable flow, and potentially dangerous density buildup. The motivating case in the original study is the Madrid Arena disaster, where poor guidance and overcrowding contributed to fatal crushing (Lopez-Carmona et al., 2020).

The system’s central idea is to divide the evacuation space into cells and assign each cell an exit recommendation encoded by color. Evacuees carry a personal or wearable device that determines the cell they currently occupy and displays the color of the exit gate they should follow. Because the assignment is updated dynamically over time, the same cell can recommend different exits at different moments as congestion, external flows, and blockages evolve. In that sense, CellEVAC is adaptive not only spatially but temporally (Lopez-Carmona et al., 2020).

A defining feature is that CellEVAC is cell-based rather than fully individualized. All pedestrians in the same cell receive the same recommendation. This simplifies sensing, communication, and control relative to per-person route planning, while still allowing online redistribution of pedestrian streams (Lopez-Carmona et al., 2020). The original paper also emphasizes that the action space is extremely large: for the Madrid Arena discretization with 42 cells and 8 exits, it states that there are $2^{126}$ control actions at each control step, making brute-force online optimization infeasible (Lopez-Carmona et al., 2020).

This places CellEVAC within a broader class of adaptive evacuation guidance systems, but with a distinctive emphasis on interpretable behavioral control. In contrast to dynamic exit-sign systems based on capacity reservation (Desmet et al., 2013), CPN-style adaptive routing (Bi et al., 2013), or individualized congestion-aware AR guidance (Zhang et al., 2020), CellEVAC’s control logic is explicitly built from a compact discrete-choice model of exit revision rather than from a purely heuristic or path-search rule.

2. Controller architecture and information flow

In the Madrid Arena case study, the ground floor is divided into 42 regular hexagonal cells of area $9\,m^2$ and width $6\,m$, chosen as a compromise between control granularity, coverage, and computational cost (Lopez-Carmona et al., 2020). The original controller architecture contains three main blocks: Pedestrian flow estimation, Control logic, and RF transmitter. The flow-estimation block periodically receives sampled images and preprocesses them to estimate pedestrian density in each cell. These densities and derived state variables are fed to the control logic, which computes the current optimal assignment of exits to cells. The RF transmitter then broadcasts 42 tuples of the form $\{Cell, Color\}$, one per cell (Lopez-Carmona et al., 2020).

The original deployment concept assumes that each cell contains an active RFID tag transmitting a cell identifier, while evacuees carry a personal device with an RFID reader and an RF receiver. By matching the detected current cell against the broadcast cell-color map, the device lights up with the color of the recommended exit (Lopez-Carmona et al., 2020). The later architecture paper specialized this into a wearable implementation based on LED wristbands and distinguished two variants. In Type A, wristbands receive both cell-node IDs and the global $\{Cell,Color\}$ table, infer the current cell from the strongest RSSI, and then map cell ID to color. In Type B, cell nodes themselves receive controller commands and rebroadcast local color information, so wristbands infer only which color applies locally rather than resolving a global map (Lopez-Carmona et al., 2020).

The update loop is quasi-discrete. In the deployment-oriented study, the controller broadcasts the 42 $\{Cell, Color\}$ tuples every 5 seconds, pedestrian position is reevaluated every 5 seconds, and active RFID tags broadcast IDs every 3–5 seconds (Lopez-Carmona et al., 2020). Operationally, the 2020 CellEVAC paper summarizes the cycle as: initialize the evacuation model and parameter set; periodically measure pedestrian positions and densities in cells and at exits; derive attributes such as distance to exits, exit congestion, and group size along paths; use the CellEVAC MLM to score each cell-exit pair; select, for each cell, the exit with the highest utility; broadcast the cell-color map; allow evacuees to update their own exit choices using a separate Pedestrians’ MLM; and then advance the microscopic simulator until the next update (Lopez-Carmona et al., 2020).

A notable design decision is that CellEVAC’s own cell-level recommendation is deterministic: for a given cell, it selects the exit with highest utility. The authors explicitly state that this deterministic choice “prevents oscillations in the decision logic of CellEVAC,” whereas individual pedestrian exit revision remains probabilistic (Lopez-Carmona et al., 2020).

3. Behavioral model and exit-choice logic

The behavioral core of CellEVAC is a Multinomial Logit Model (MLM) derived from random utility theory. The paper uses the standard decomposition

$U_{pj}=V_{pj}+\varepsilon_{pj},$

with systematic utility

$V_{pj}=\sum_k \beta_{jk}x_{pjk},$

and, under IID Gumbel errors, the standard logit choice probability

$$P_{pj}=\frac{\exp(V_{pj})}{\sum_{E_i\in E(p)}\exp(V_{pi})}. \tag{1}$$

This probability model is used for pedestrian exit revision, while CellEVAC itself uses the same utility logic but applies deterministic argmax selection at the cell level (Lopez-Carmona et al., 2020).

The authors deliberately keep the utility specification compact and behaviorally interpretable. The six modeled attributes are Distance, Exit width, Group size along the path, Congestion at exits ($EXCON$), Personal tendency to maintain the current exit decision, and System indication from CellEVAC. The printed utility equation in the paper is partially corrupted, but the accompanying prose makes the intended structure explicit: $9\,m^2$0 Distance is normalized by the maximum scenario distance, with $9\,m^2$1 expected to favor nearer exits. Exit width is normalized by maximum width, with $9\,m^2$2 favoring wider gates. The $9\,m^2$3 term is a path-congestion indicator computed from the cell partition by summing pedestrians in cells closer to each candidate exit. The $9\,m^2$4 term is normalized by the exit’s critical density,

$9\,m^2$5

so values above 1 indicate that the exit is operating in a highly congested regime (Lopez-Carmona et al., 2020).

Two binary terms govern revision dynamics. $9\,m^2$6 if exit $9\,m^2$7 is the pedestrian’s current choice and 0 otherwise, thereby modeling inertia. Its coefficient is time-dependent: $9\,m^2$8 At the start of evacuation, $9\,m^2$9, so pedestrians are relatively willing to revise; as the crowd thins, $6\,m$0 rises toward $6\,m$1, increasing commitment to prior choices. $6\,m$2 if CellEVAC recommends exit $6\,m$3 to pedestrian $6\,m$4, and 0 otherwise, so adaptive instructions increase the attractiveness of the recommended exit in proportion to $6\,m$5 (Lopez-Carmona et al., 2020).

Compliance is modeled as a mixture of populations. A certain fraction of evacuees is assumed to follow CellEVAC throughout the evacuation, while the remainder ignores the system and follows the unguided STANDARD pedestrian model. At the extremes, if $6\,m$6, pedestrians do not respond to the system, whereas if all coefficients are zero except $6\,m$7, they strictly follow CellEVAC (Lopez-Carmona et al., 2020). The later deployment paper reformulated the controller-side inertia term as $6\,m$8 with coefficient $6\,m$9, but the operational meaning is the same: recommendation stability is weak early in evacuation and stronger later (Lopez-Carmona et al., 2020).

The original paper reports optimized coefficient sets for three configurations. For STANDARD, $\{Cell, Color\}$0, $\{Cell, Color\}$1, $\{Cell, Color\}$2, $\{Cell, Color\}$3, and $\{Cell, Color\}$4. For OPTIMAL individual behavior, $\{Cell, Color\}$5, $\{Cell, Color\}$6, $\{Cell, Color\}$7, $\{Cell, Color\}$8, and $\{Cell, Color\}$9. For CellEVAC, $\{Cell,Color\}$0, $\{Cell,Color\}$1, $\{Cell,Color\}$2, $\{Cell,Color\}$3, and $\{Cell,Color\}$4. The authors interpret the sign reversal in $\{Cell,Color\}$5 as evidence that once cell-based recommendations already induce grouping, the controller must actively discourage excessive herd-like concentration (Lopez-Carmona et al., 2020).

4. Simulation, optimization, and explicit safety modeling

CellEVAC is not only a communication concept but a simulation-optimization framework. Microscopic pedestrian motion is modeled using the classical Social Force Model (SFM) through AnyLogic’s pedestrian library, while decision/control logic is implemented in Matlab and optimized with Tabu Search (TS) via AnyLogic’s OptQuest engine (Lopez-Carmona et al., 2020). The optimization targets three objects: the Pedestrians’ MLM, the CellEVAC MLM, and, for comparison, a Cartesian Genetic Programming (CGP) controller.

For MLM-based optimization, each candidate is a set of $\{Cell,Color\}$6 coefficients. The objective is

$\{Cell,Color\}$7

where $\{Cell,Color\}$8 is a non-positive safety measure, so minimizing $\{Cell,Color\}$9 simultaneously reduces evacuation time and improves safety. Because early CGP solutions often failed to evacuate all pedestrians, the comparison controller required a modified objective: $\{Cell, Color\}$0 The original optimization used 200 iterations, each parameter setting was evaluated by a replication algorithm with 3 to 10 stochastic simulation replications, and a hard simulation stop limit of 15 minutes evacuation time was imposed to avoid wasting search effort on hopeless solutions (Lopez-Carmona et al., 2020).

A distinctive feature of CellEVAC is that safety is an explicit optimization target, not merely an indirect byproduct of faster evacuation. The authors model exit-gate dynamics using pedestrian Macroscopic Fundamental Diagrams (MFDs) derived from microscopic simulation under stressful inflow patterns. For each exit, they identify three characteristic densities: $\{Cell, Color\}$1, the density at maximum capacity and end of free flow; $\{Cell, Color\}$2, the stable overcrowded regime after shockwave formation; and $\{Cell, Color\}$3, severe accumulation under blocking. The MFD curves are fitted using a sixth-order polynomial with robust bisquare weighting (Lopez-Carmona et al., 2020).

The per-exit safety value is defined through normalized density exceedance above a safety threshold. The printed equations are partially corrupted, but the intended construction is explicit in the text: $\{Cell, Color\}$4 with

$\{Cell, Color\}$5

and $\{Cell, Color\}$6. The interpretation is operationally clear: if density at an exit always stays below $\{Cell, Color\}$7, then $\{Cell, Color\}$8; if the exit effectively locks and density approaches $\{Cell, Color\}$9, then $U_{pj}=V_{pj}+\varepsilon_{pj},$0. System-wide safety is then summarized by

$U_{pj}=V_{pj}+\varepsilon_{pj},$1

and safety imbalance across exits by

$U_{pj}=V_{pj}+\varepsilon_{pj},$2

This formulation makes unstable, highly variable, or unevenly distributed compression at exits directly visible in the optimization target (Lopez-Carmona et al., 2020).

5. Madrid Arena case study and empirical findings

The main case study is the Madrid Arena ground floor, a venue with area $U_{pj}=V_{pj}+\varepsilon_{pj},$3, maximum capacity 3,400 spectators, 8 exit gates, and exit widths from 2.5 m to 6 m (Lopez-Carmona et al., 2020). To model interactions with upper floors, additional pedestrian inflows are injected at exits 1, 2, 3, 4, and 6. In the principal complex-flow experiments, at each simulation iteration three exits are chosen randomly; two receive incoming flows of $U_{pj}=V_{pj}+\varepsilon_{pj},$4; and the third is blocked. All evacuation experiments use 3400 ground-floor evacuees with preferred speeds uniformly distributed between $U_{pj}=V_{pj}+\varepsilon_{pj},$5 and $U_{pj}=V_{pj}+\varepsilon_{pj},$6 (Lopez-Carmona et al., 2020).

The central quantitative finding is that CellEVAC outperforms unguided evacuations, and the improvement becomes exponential as interactions become more complex. The paper does not provide a separate exponential law, but uses this wording to characterize the empirical growth in benefit under stronger interference from blocked exits and external flows (Lopez-Carmona et al., 2020). Relative to the idealized OPTIMAL individual behavior benchmark, CellEVAC achieves a median total evacuation time of around 8 minutes, only about 30 seconds above OPTIMAL, indicating that a practical cell-based guidance policy can approach the benchmark of optimized individual adaptation (Lopez-Carmona et al., 2020).

The comparison with CGP is also central. CellEVAC performs better overall in terms of safety, evacuation time, and number of revisions of exit-choice decisions. The authors further report that CGP yields less natural pedestrian reactions and movements, with larger oscillations and a right-skewed distribution of decision changes. Their interpretation is that a controller built on a behavioral model is more likely to generate stable and plausible pedestrian responses than a controller built as a directly evolved opaque heuristic (Lopez-Carmona et al., 2020).

Decision-change dynamics are treated as a behavioral realism issue rather than a cosmetic side effect. In a sensitivity study using only distance, as $U_{pj}=V_{pj}+\varepsilon_{pj},$7 became less strongly negative, uncertainty increased, the number of decision changes rose exponentially, evacuation time worsened, and in some cases pedestrians effectively failed to evacuate. When the $U_{pj}=V_{pj}+\varepsilon_{pj},$8 term was introduced, behavior became much more stable. For example, with $U_{pj}=V_{pj}+\varepsilon_{pj},$9, the median evacuation time in the no-external-flow study dropped to about 4.8 minutes, compared with 6–12 minutes without the personal inertia term (Lopez-Carmona et al., 2020).

Compliance experiments further show that CellEVAC remains beneficial even when only part of the population follows the system. The paper varies compliance from 0% to 100% in 20% increments and reports that the proposal has a positive influence even for a 40% compliance rate. In those experiments, evacuation time reaches its best value at 60% compliance, while average safety, safety variance, and number of decision changes all improve approximately linearly with compliance. The authors suggest that imitation effects in the crowd may amplify the influence of the guided subgroup (Lopez-Carmona et al., 2020).

6. Deployment realism, uncertainty, limitations, and research context

The second CellEVAC paper moves from idealized control to a more explicit cyber-physical deployment architecture based on a controller node, a cell-node network with active RFID devices, and radio-controlled LED wristbands (Lopez-Carmona et al., 2020). It models cell-level localization by strongest-signal selection under a log-normal RSSI propagation law. In simplified form, for the chosen simulation parameters,

$V_{pj}=\sum_k \beta_{jk}x_{pjk},$0

where $V_{pj}=\sum_k \beta_{jk}x_{pjk},$1 is a Gaussian random variable with standard deviation $V_{pj}=\sum_k \beta_{jk}x_{pjk},$2, used as the uncertainty parameter (Lopez-Carmona et al., 2020).

This deployment study shows that localization quality is the main practical bottleneck. When the control logic optimized for zero uncertainty is used, evacuation time begins to increase roughly linearly for $V_{pj}=\sum_k \beta_{jk}x_{pjk},$3 dB in no-external-flow scenarios, decision changes increase sharply, and average safety deteriorates. The bottom-line recommendation is explicit: CellEVAC is operationally useful only when RSSI random variation is roughly below 10 dB, and “cannot be applied in a real environment if the standard deviation of the RSSI values is greater than 10 dB” (Lopez-Carmona et al., 2020). Re-optimizing the MLM under expected uncertainty gives only marginal improvements below 20 dB and no meaningful advantage in the regime where the system is actually usable, so the paper concludes that improving the positioning subsystem matters more than retuning the controller (Lopez-Carmona et al., 2020).

Several limitations are acknowledged in the original CellEVAC study. The safety threshold

$V_{pj}=\sum_k \beta_{jk}x_{pjk},$4

is chosen for comparison purposes rather than empirically validated as a universal safety boundary. Safety is assessed only at exit gates, not across the full facility. The Madrid Arena geometry is comparatively simple: exits are directly visible, there are no obstacles affecting line of sight, and cell sizes are homogeneous. The paper also assumes a functioning cell-level positioning and communication infrastructure, but does not experimentally validate the sensing stack or deployment economics. Finally, the behavioral model is intentionally compact, with homogeneous coefficients within each population and no richer stochastic non-compliance model beyond the guided-versus-unguided mixture assumption (Lopez-Carmona et al., 2020).

Within the wider evacuation-guidance literature, CellEVAC can be read as an overview of adaptive control and behavioral modeling. Dynamic-sign systems based on future capacity reservation seek flow balancing through local directional actuation rather than wearable individualized signals (Desmet et al., 2013). CPN-based evacuation routing emphasizes distributed adaptive path discovery and class-specific QoS objectives, but not the same cell-based common recommendation mechanism (Bi et al., 2013). AR-based congestion-aware routing computes individualized multi-exit paths from real-time occupancy maps, whereas CellEVAC trades some personalization for simpler cellwise coordination (Zhang et al., 2020). A plausible implication is that CellEVAC’s main methodological contribution lies in showing that adaptive crowd guidance can be formulated as a behaviorally optimized, cell-level recommendation problem rather than only as a shortest-path, signal-scheduling, or heuristic-evolution problem.