Data-driven framework for the adaptive exit selection problem in pedestrian flow: Visual information based heuristics approach

https://doi.org/10.1016/j.physa.2021.126289Get rights and content

Abstract

Pedestrian behavior during evacuation has been formulated using various arbitrary microscopic methods to investigate the performance of crowd dynamics while their custom rules result in low visual realism in simulation due to the complexity of intrinsic decision logic of human. Statistical analysis is an effective way to reveal the motion pattern and path planning behavior of pedestrians whose main idea is to approach the trajectory and social attributes data of pedestrians extracted from evacuation drills as much as possible. In this study, we present a data-driven based microscopic pedestrian-simulation model with continuous-space representation to explore the potential of integrating empirical analysis into crowd simulation to enhance the authenticity of decision making. This method extracts the pedestrian’s decision mode and smoothly applies it in the crowd dynamics model. Instead of navigating agents by arbitrary regulations, the desired direction of pedestrians during the motion is arranged by machine learning (ML) algorithms. The path decision module trained with actual pedestrian data improves the compatibility of the model in the application of various spatial scenarios and no longer suffers from tedious parameter fine-tuning work. To completely describe the information precepted by pedestrians, a polygon segmentation module is developed to divide the visual field of pedestrians and identify the mutual visibility among them. This module filters out the information that can be perceived by pedestrians in real situations, thereby bridges the gap between statistical analysis and numerical simulation methods. We compare different ML approaches for route-choice behavior prediction and discuss the relative importance of its influencing variables under different scenarios. Inferring the perception of social interactions from disaggregate choice data, the scope of effectiveness of conformity behavior and crowd-aversion are also discussed. The simulation results are compared with experimental data, illustrating the model’s capability to accurately reproduce the observed flow motion in various scenarios with moderate modification in physical environment initialization.

Introduction

Crowd dynamics is a challenging issue due to the inherent complexity of human behaviors, the variety of geometrical layouts, the context-dependency of human behaviors etc. Revealing the variable patterns of movement characteristics can place occupants as beneficiaries in facilitate building safety design and evacuation management and planning. Irregularity of the interior space of the building and uncertainty in utility evaluation on different routes directly determines pedestrian’s travel time and trajectory, and also affects the evacuation efficiency of pedestrian flows.

Pedestrian behavior during an evacuation can be viewed as the result of hierarchical decision-making processes, i.e. strategic level (departure time choice and activity pattern choice); tactical level (activity scheduling, activity area choice, and route choice); operational level (direction and velocity) [1], [2], [3], [4], [5], [6], [7], [8], [9], [10], [11]. Crowd dynamics simulation is a powerful tool to evaluate of the service level and safety performance prior to guidance system implementation and facility design. To characterize pedestrian behavior pattern and evacuation processes, a wide range of physical models were developed, such as Cellular Automata (CA) model [12], [13], [14], [15], [16], Social Force (SF) model [17], [18], [19], and Lattice Gas (LG) based model [20], [21], [22]. These models provide a straight mathematical framework, while relating model parameters based on logical principles. CA and LG adopt discrete modeling method in which the space is divided into regular polygons. Pedestrians choose the purpose of movement by quantifying the utility of the grid corresponding to each position. The difference between these two methods is that in the LG model, there may be multiple individuals in the same grid at the same time interval, while in the CA model, a space can only be occupied by one pedestrian. The SF model is a continuous based model that simulates pedestrians as particles that receive various forces, and these forces drive pedestrians to move in space. Compared with the previous two models, it is more compatible with space, but it also has some limitations, such as the inability to simulate discontinuous behaviors such as pedestrians’ sudden stops and right-angle turns.

All models quantify various features in the space, thereby driving pedestrians to make optimal decisions from the optional action space. To arrange the setting of behavioral regulations in the simulation models, analysis of observational or experimental data appears to be a prerequisite. By taking into account those factors noticed in site investigation such as the availability of guidance system [23], [24], evacuees familiarity with space layout [25], degree of collision avoidance [3], [26], aversion of congestion [27], [28] etc, evacuation models were subsequently improved and polished.

In recent years, the fineness of the model gradually deepens in both kinetics and psychological description, therefore the compatible scenes of simulation models are becoming more and more extensive. Evacuation scenes in some typical geometrical layouts, such as corners [29], corridors [30], [31], high rise building [32], [33], [34], staircases [35], [36], T-shaped channels [37], have been reproduced through experiments and simulations. The unique features of pedestrian flow result in these special geometrical space are compared with experimental data. Much effort has been put in study of information perception and decision-making in evacuation processes, either. Simulation models use bi-layer guidance [3], [7], [8], [38], Eikonal equations [39], [40], Fractal Brownian motion [41] and other methods to include more evaluation factors in the description of the space potential field. The gradually complete decision-making mechanism enables many special behaviors, such as queuing, following, path planning, etc., to be reproduced in the model and related optimization schemes can be obtained through sensitivity analysis.

However, most modeling studies have not paid enough attention to the decision-making psychology of simulated pedestrians and the preferences of various factors in real decision-making. To say the least, there is no definite quantitative method for the information accessed by agents in the model, and the perception of these information is not reasonably filtered and screened. Shortcomings of the model based-on arbitrary scenarios and regulations are mainly reflected in the difficulty in determining reasonable parameters and calculation methods. First, it is assumed that the pedestrians share the same information field, no matter where they are scattered. Significantly, researchers have ignored the effects of the visibility diversity on pedestrian decision-making behavior. Simplifying the transmission of information to the diffusion behavior among potential field cells will naturally hardly reflect the obstruction of the line of sight by obstacles. At the same time, this mode of intercellular information transmission will inevitably produce perceptual hysteresis and fails to simulate the immediacy of vision. Second, the interaction between pedestrians and environment can only be realized through the information field on the basis of some assumptions. In other words, there is currently no clear and effective plan for the definition and calculation method of the utility of each position in the site space and pedestrian interaction.

Taking a step back, this kind of modeler-defined utilities do not link to the any scalar in practical scene or empirical analysis. It is still lack of a precise method to measure the imaginary potential and the interaction force on psychological level [42], [43], let along calibration. As important indicators in the model, how to convert them to generalized travel cost still depends on empirical or intuitive evaluation rules. Third, the weight of various influencing factors in pedestrian decision-making is still unclear. For example, in some model based on potential fields, the degree of congestion, distance, and the attraction of the leader are exponentially added. The repulsion among pedestrians is calculated using a method similar to the principle of interaction between charges [43], [44]. Even in the empirical studies, there is no proper conclusion about these two psychological quantitative methods that have opposite effects. Some studies suggested that the pedestrians are inclined to follow majority in path planning decision [45], [46], [47], and a few other studies found that pedestrian act congestion aversion in this utility evaluation [48], [49], [50]. Therefore, it is inappropriate to simply superimpose psychological effects of conformity behavior and congestion aversion with the same quantitative method, which causes confusion in model formulation. These assumptions limit the expressive ability of the model, and more adjustments must be made to accommodate these assumptions.

Given the rapid development of affordable robots with embedded sensing and computation capabilities, data-driven crowd model has been studied using various methods as an emerging classical subject in complexity science to enhance the visual realism of simulation [51], [52], [53], [54], [55], [56], [57], [58], [59]. Methods and techniques on cluster, classify and regression with no explicit a prior definition of model regulation, and these methods are not subject to parametric limitations. Research work in trajectory prediction have proposed several data-driven methods to address the interaction among agents [60], [61], [62]. Charalambous and Chrysanthou [63] extracted the data of the target scene and connected the input instance to a perceptual action graph for the simulation. Liu et al. [64] trained the model by the trajectory data from video record, and stored them in path sets to guide crowd simulation. Based on their works, Yao et al. [65], [66] improve the scene adaptability of model by divide the crowd features into physical and social properties. Establish the residual network by the crowd properties as parameters, this approach allows the scene-independent prediction to fit the movement behavior more accurately. However, the goal of these studies biased to describe robot motion is to obtain a stable and reliable collision avoidance method rather than a route selection strategy for pedestrian [67], [68]. As an NP-hard problem, multi-agent path finding (MAPF) does have some approximate solution based on optimal method. The primary challenge here is that the environment changes dynamically along the progress of simulation, where the influence of agent behavior at current on others in the future will be difficult to quantify.

Several approaches using deep neural networks in reinforcement learning have recently been proposed and gain remarkable improvement in crowd simulation, even though some of them show plausible results. Nonetheless, the rules of information transmission is a limitation that hinders the further expansion of such methods to crowd dynamics simulation. The training or policy execution process of many existing methods rely on explicit communication between agents to share observations or selected actions, which is inconsistent with the real situation in pedestrian flow [69], [70], [71]. Although data analysis from experiments has revealed many behavior patterns of pedestrians, these results are difficult to actually apply in the model. One of the reasons for this is that most models do not include a suitable module to describe pedestrians’ perception of information, which also hinder the application of the model obtained through experimental fitting to crowd dynamics simulation. Of course computational simulation do not faithfully incorporate many of the visual, auditory, olfactory and somatosensory inputs that humans experience in real life, but the introduction of visual information can always improve the compatibility between the information transmission section and the decision-making model obtained in the experiment.

Pedestrian’s visual field constantly changes as they move around a room, due to the obstruction of structural components or large furniture. Changes in position simultaneously update optional route sets and pedestrians in view. Recently, some researchers have linked visual information with pedestrian’s movement. The core module, the expression of the scope of vision, mainly has two types of methods to formulate. The first one is discrete grid based method, which is similar to the division method of the potential field method in crowd dynamics simulation. The presetting static representation marks the objects that pedestrians can see within each position by pointers among grids. Proposing an individual navigation model that, with the aid of an ‘exosomatic visual architecture’, generates a visibility graph by meshing the facility into small squares [72], [73]. In this case, the model cannot get satisfactorily accurate results in certain environments, such as irregular space with non-right-angled components. The second method is to dynamically define a fan-shaped area for each agent to represent its visual field [3], [44], [74], [75], [76]. Then the angle and radius of this sector are often assumed based on intuition and experience. Due to its small radius, the field of vision of this type of model is only used to collect special information but the pedestrians factors in the visual field, such as flow velocity and congestion level. In other words, such short-sighted agent cannot conduct path planning and global information gathering behavior at the tactical level. In this way, it is a problem worth exploring to accurately describe the pedestrian’s field of view in the space with internal obstacles and accurately extract visible information. Using visual information to replace the potential field representation can directly extract various features of pedestrians and avoid inappropriate quantification in route utility evaluation.

The aim of this research is to propose a data-driven route-choice module and investigate its appropriateness of use alongside microscopic crowd dynamic model. By introducing the visual field division module to bridge the gap between experimental analysis and simulation method, the authenticity and adaptability of model can be improved. The model filters out information that pedestrians cannot perceive, and uses non-parametric methods to conduct route-choice, so that avoid defining arbitrary regulations with confusing physical meaning to transform pedestrian characteristics into travel cost. Getting rid of assumptions and simplifications, this bi-layer model can be applied to various irregular spatial space and visibility condition, and the characteristics of pedestrians, such as movement pattern, decision-making behaviors, interactive phenomena, and visual information perception, can be described in detail.

This study is organized as follows. In Section 2, we describe the experimental setup and data collection methods and the general classification algorithms used in path-planning. In sequence, the hybrid modeling framework with data-driven routing and activity implementation of pedestrian is presented. An algorithm for extracting visual information for pedestrian is developed to coordinate the operation of the two layers of microscopic crowd dynamics model. Detailed analysis of route-choice prediction and evacuation dynamics are conducted in Section 3, followed by an elucidation of discussion of this study (Section 4). Conclusion and further application scenario are presented to explore the potential of our model in Section 5.

Section snippets

Hybrid modeling framework

The training data used in the model in this work is extracted from egress drills. In this scenario, volunteers have peaceful mentality and a clear decision-making logic without a strong sense of competition. Some exits in the space are visible, and other exits are blocked by obstacles. Compared with the multi-directional pedestrian flow in public places, rational pedestrians tend to move to a few fixed destinations (exits) in an evacuation situation and care more about time cost instead of

Comparison among machine learning exit choice prediction approaches

As discussed earlier, the dataset containing 4 variables that obtained from experiments by extract the record of pedestrian route-choice behavior, and the summary statistics of these variables can be seen in Haghani and Sarvi [84]. There are in total 21659 selection data are collected from a total of 1204 trials, and each choice situation is a unique combination of four design variables. Those variables include (1) the level of congestion of occupants around each exit door (CONG); (2) the

Discussion

Our main contribution can be summarized as an attempt to establish a connection between two isolated research methods, experimental analysis and numerical simulation, improve the authenticity of the model, and broadening the scope of application of the simulation method based on this framework. In general, parameter-based models proposed by previous studies often require massive sensitivity analysis to debug key attribute parameters in every single case. Different with other models that can

Conclusion

Crowd dynamics behavior in public space that incorporates the assignment of exit distribution and visual information is essential concern in building design for serve residents. However, the difficulty is that it consists of multiple factors with vague physical relationship among each other including crowd density and flow velocity, conformity behavior and congestion aversion, even the description of crowd density does not yet have a widely recognized physical unit. To circumvent those problems

CRediT authorship contribution statement

Zi-Xuan Zhou: Writing – original draft, Conceptualization. Wataru Nakanishi: Writing – review & editing, Conceptualization. Yasuo Asakura: Writing – review & editing, Conceptualization.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgments

The authors are grateful to Ms. Xuefei Li for her helpful comments. Special thanks go to the anonymous reviewers for their great suggestions.

References (96)

  • BursteddeC. et al.

    Simulation of pedestrian dynamics using a two-dimensional cellular automaton

    Physica A

    (2001)
  • FlötterödG. et al.

    Bidirectional pedestrian fundamental diagram

    Transp. Res. B

    (2015)
  • FuZ. et al.

    A fine discrete field cellular automaton for pedestrian dynamics integrating pedestrian heterogeneity, anisotropy and time-dependent characteristics

    Transp. Res. C

    (2018)
  • PorterE. et al.

    Pedestrian dynamics at transit stations: an integrated pedestrian flow modeling approach

    Transp. A: Transp. Sci.

    (2018)
  • SticcoI. et al.

    Social force model parameter testing and optimization using a high stress real-life situation

    Physica A

    (2021)
  • IsobeM. et al.

    Experiment and simulation of pedestrian counter flow

    Physica A

    (2004)
  • WangX. et al.

    Information guiding effect of evacuation assistants in a two-channel segregation process using multi-information communication field model

    Saf. Sci.

    (2016)
  • CaoS. et al.

    Modeling pedestrian evacuation with guiders based on a multi-grid model

    Phys. Lett. A

    (2016)
  • LiD. et al.

    Behavioral effect on pedestrian evacuation simulation using cellular automata

    Saf. Sci.

    (2015)
  • LachapelleA. et al.

    On a mean field game approach modeling congestion and aversion in pedestrian crowds

    Transp. Res. B

    (2011)
  • LiS. et al.

    Block-based floor field model for pedestrian’s walking through corner

    Physica A

    (2015)
  • HeliövaaraS. et al.

    Pedestrian behavior and exit selection in evacuation of a corridor - an experimental study

    Saf. Sci.

    (2012)
  • PelechanoN. et al.

    Evacuation simulation models: Challenges in modeling high rise building evacuation with cellular automata approaches

    Autom. Constr.

    (2008)
  • HuoF. et al.

    Experimental study on characteristics of pedestrian evacuation on stairs in a high-rise building

    Saf. Sci.

    (2016)
  • LiS. et al.

    Occupant evacuation and casualty estimation in a building under earthquake using cellular automata

    Physica A

    (2015)
  • ShahhoseiniZ. et al.

    Pedestrian crowd flows in shared spaces: Investigating the impact of geometry based on micro and macro scale measures

    Transp. Res. B

    (2019)
  • AbdelghanyA. et al.

    A hybrid simulation-assignment modeling framework for crowd dynamics in large-scale pedestrian facilities

    Transp. Res. A

    (2016)
  • HughesR.L.

    A continuum theory for the flow of pedestrians

    Transp. Res. B

    (2002)
  • GuoR.-Y. et al.

    Formulation of pedestrian movement in microscopic models with continuous space representation

    Transp. Res. C

    (2012)
  • HuJ. et al.

    Study on queueing behavior in pedestrian evacuation by extended cellular automata model

    Physica A

    (2018)
  • SteffenB. et al.

    Methods for measuring pedestrian density, flow, speed and direction with minimal scatter

    Physica A

    (2010)
  • XiaoY. et al.

    A pedestrian flow model considering the impact of local density: Voronoi diagram based heuristics approach

    Transp. Res. C

    (2016)
  • FangJ. et al.

    Leader-follower model for agent based simulation of social collective behavior during egress

    Saf. Sci.

    (2016)
  • KinatederM. et al.

    Social influence in a virtual tunnel fire - influence of conflicting information on evacuation behavior

    Applied Ergon.

    (2014)
  • LinJ. et al.

    Do people follow the crowd in building emergency evacuation? A cross-cultural immersive virtual reality-based study

    Adv. Eng. Inf.

    (2020)
  • HaghaniM. et al.

    How perception of peer behaviour influences escape decision making: The role of individual differences

    J. Environ. Psychol.

    (2017)
  • HaghaniM. et al.

    Following the crowd or avoiding it? Empirical investigation of imitative behaviour in emergency escape of human crowds

    Anim. Behav.

    (2017)
  • HaghaniM. et al.

    ‘Herding’ in direction choice-making during collective escape of crowds: How likely is it and what moderates it?

    Saf. Sci.

    (2019)
  • LovreglioR. et al.

    A mixed logit model for predicting exit choice during building evacuations

    Transp. Res. A

    (2016)
  • MaL. et al.

    The analysis on the desired speed in social force model using a data driven approach

    Physica A

    (2019)
  • MartinR.F. et al.

    Data-driven simulation of pedestrian collision avoidance with a nonparametric neural network

    Neurocomputing

    (2020)
  • LiuB. et al.

    A social force evacuation model driven by video data

    Simul. Model. Pract. Theory

    (2018)
  • YaoZ. et al.

    Data-driven crowd evacuation: A reinforcement learning method

    Neurocomputing

    (2019)
  • YaoZ. et al.

    Learning crowd behavior from real data: A residual network method for crowd simulation

    Neurocomputing

    (2020)
  • WangK. et al.

    A machine learning based study on pedestrian movement dynamics under emergency evacuation

    Fire Saf. J.

    (2019)
  • WangK. et al.

    Influence of human-obstacle interaction on evacuation from classrooms

    Autom. Constr.

    (2020)
  • ZengW. et al.

    Application of social force model to pedestrian behavior analysis at signalized crosswalk

    Transp. Res. C

    (2014)
  • WangW. et al.

    Microscopic modeling of pedestrian movement behavior: Interacting with visual attractors in the environment

    Transp. Res. C

    (2014)
  • Cited by (6)

    View full text