(C) PLOS One
This story was originally published by PLOS One and is unaltered.
. . . . . . . . . .



Identification of social groups and waiting pedestrians at railway platforms using trajectory data [1]

['Mira Küpper', 'School Of Architecture', 'Civil Engineering', 'University Of Wuppertal', 'Wuppertal', 'Armin Seyfried', 'Institute For Advanced Simulation', 'Forschungszentrum Jülich', 'Jülich']

Date: 2023-05

Abstract To investigate the impact of social groups on waiting behaviour of passengers at railway platforms a method to identify social groups through the monitoring of distances between pedestrians and the stability of those distances over time is introduced. The method allows the recognition of groups using trajectories only and thus opens up the possibility of studying crowds in public places without constrains caused by privacy protection issues. Trajectories from a railway platform in Switzerland were used to analyse the waiting behaviour of passengers in dependence of waiting time as well as the size of social groups. The analysis of the trajectories shows that the portion of passengers travelling in groups reaches up to 10% during the week and increases to 20% on the weekends. 60% of the groups were pairs, larger groups were less frequent. With increasing group size, the mean speed of the members decreases. Individuals and pairs often choose waiting spots at the sides of the stairs and in vicinity of obstacles, while larger groups wait close to the platform entries. The results indicate that passengers choose waiting places according to the following criteria and ranking: shortest ways, direction of the next intended action, undisturbed places and ensured communication. While individual passengers often wait in places where they are undisturbed and do not hinder others, the dominating comfort criterion for groups is to ensure communication. The results regarding space requirements of waiting passengers could be used for different applications. E.g. to enhance the level of service concept assessing the comfort of different types of users, to avoid temporary bottlenecks to improve the boarding and alighting process or to increase the robustness of the performance of railway platforms during peak loads by optimising the pedestrian distribution.

Citation: Küpper M, Seyfried A (2023) Identification of social groups and waiting pedestrians at railway platforms using trajectory data. PLoS ONE 18(3): e0282526. https://doi.org/10.1371/journal.pone.0282526 Editor: Iman Aghayan, Shahrood University of Technology, IRAN, ISLAMIC REPUBLIC OF Received: October 20, 2022; Accepted: February 16, 2023; Published: March 15, 2023 Copyright: © 2023 Küpper, Seyfried. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Data Availability: The trajectory data used for this study are the property of Swiss Federal Railways (SBB AG). The data is available upon request addressed to the Swiss Federal Railways (Contact: [email protected]) and after signing a confidentiality agreement (as it was done by the authors). Funding: This study is a part of the project “CroMa – Crowd Management in Verkehrsinfrastrukturen” which is funded by the German Federal Ministry of Education and Research (BMBF) under grant number 13N14530. The funders had no role in the study design, data collection and analysis, decision to publish or the preparation of the manuscript. There was no additional external funding received for this study. Competing interests: The authors have declared that no competing interests exist.

Introduction The movement of pedestrians is studied in different situations, e.g. evacuations of buildings or large-scale events, and was analysed in both field and laboratory experiments; for an overview see [1–3] and the conference series [4, 5]. Most previous studies focused on characteristics of pedestrian flows and the parameters that influence their movement. The key concepts for evaluating a facility regarding pedestrian traffic include the fundamental diagram and the Level of Service (LOS) concept [6, 7]. The fundamental diagram is used to estimate the capacity of pedestrian facilities and whether heavy congestion occur. The Level of Service concept allows to rate the comfort at a certain density. While those concepts provide information on the comfort pedestrians feel and whether the pedestrian load can lead to dangerous situations, they were designed for environments in which movements occur. Waiting pedestrians were solely considered while standing in queues, as for example in the LOS presented in [8]; but obviously the interaction between moving and waiting and therewith standing pedestrians could have a great influence on the dynamic and must be considered in dependence of the context. Larger social groups in particular can be an impactful obstacle in pedestrian flows. This is for example the case at train station platforms, where both moving and waiting passengers are present simultaneously. The first part of this introduction focuses on research on waiting behaviour at train platforms, and the second part addresses aspects of the research on social groups. Only limited research has been published to date about pedestrian waiting behaviour and the factors that influence how and where pedestrians wait at railway platforms and how these waiting positions influence the pedestrians moving on the platform. Previous studies on this topic revealed that pedestrians tend to cluster in the vicinity of the platform entries [9–12] and around obstacles [13]. Seating arrangements are frequently used as waiting place, beginning with the ones closest to the platform entries [9]. The ticket machines were found to lead to crowding and congestion [14, 15], especially if placed close to stairs. Even if the stopping positions of trains are indicated on the overhead signals, passengers tend not walk to the farther ends of the platform as they do not trust the information or are unaware of its existence [9]. Reference [16] state that a confusing station layout leads to longer passenger waiting times, as pedestrians tend to arrive to such stations earlier in order to ensure they find their way. Moreover, the type of passengers using the train station platforms has an impact on the distribution and waiting behaviour. Depending on the purpose of the journey, passengers carry different amounts of luggage and possess varying degrees of familiarity with the environment. Passengers carrying luggage e.g., become important when the vertical or horizontal gaps between platform and train are larger [17], as those will increase the boarding time. Commuters often develop individual strategies [18] to minimise travel times and therefore for example wait in places where the train car that provides the shortest way at the desired destination is expected to arrive. This literature review highlights the previous studies on waiting pedestrians in the context of railway platforms. Those studies made no differentiation whether the pedestrians were individual persons travelling alone or members of social groups. Such a distinction however is necessary in order to interpret the findings and to respect the characteristics of individuals and social groups. This differentiation can help to sharpen the findings obtained on passenger’s waiting behaviour. Instead of considering a pedestrian crowd as consisting solely of a certain number of separate individuals who have no social relation, a crowd is rather to be understood as a gathering of individuals and small groups that are at the same place at the same time [19–21]. The dynamics of inter-group behaviour are proposed as social identity theory and self-categorisation theory by [22, 23]. The effect of group behaviour on pedestrian movements has become a growing research area. The following will present the main findings of previous studies, which reveal differences between (moving) groups and individual persons in public environments. Depending on the group size and the density conditions social groups are expected to walk in specific manners: small groups of two to three members tend to walk side by side in low density environments [24] and form lines perpendicular to the groups walking direction, causing such groups to occupy a large area. With increasing density and therewith limited available space, groups adapt their walking behaviour and move in “V” or “U”-like formations [24–26]. Usually the central pedestrian in those configurations walks in the rear, ensuring the groups communication. Large groups split up into smaller subgroups, since communication with all group members becomes impossible [24]. Groups are slower than individuals and with increasing group size the velocity of the group members reduces; this was observed regardless of the density [26–28]. However, in high density conditions the velocity differences between members of social groups and individual persons become smaller, as groups give up their social interaction in favour of collision avoidance and start walking in single file [28]. [29, 30] analysed group sizes of free-forming small groups and found that each group size is less frequent than the next smaller group size. In field observations social groups can be identified by the relation of their members. This relation is indicated by communication that is composed of oral and non-verbal elements such as gestures, body language and eye contact. Recent technical achievements in data collection (c.f. [31, 32]) enabled the collection of large trajectory data sets, which prompted the development of methods to analyse the movement of social groups. For example [24, 26–28, 33–35] use video recordings and trajectory data from public spaces to analyse and develop dynamical models for the movement of pedestrians in groups. The combination of video recordings and trajectory data offers the possibility to generate an annotated data set, in which information extracted from the videos, e.g., the visual identification of socially related pedestrians, can be transferred to the trajectories. Such an approach enables the analysis of the data of known social group members with respect to interpersonal distances, motion direction or angles between the group members velocity vectors. However, all these studies focus on pedestrians that are walking. A method using trajectory data applicable for waiting pedestrians was proposed by [36]. Social groups at train station platforms were identified based on their space and time relation. Pedestrians who showed a pairwise distance of below 1.5 m for 90% and a distance below 1 m for 40% of the time they spend at the platform were considered to be a social group (see also section Methodology). The study was performed with data collected in the first phase of the COVID-19 pandemic in 2020 and analysed with respect to contact tracing and distancing rules. Up to 70% of pedestrians moving in urban environments can be assigned to social groups [24, 37], during events (such as sport events or public celebrations) the portion can be even higher [25]. It than follows that the presence of social groups impacts the dynamics of pedestrian flows. Simulations indicate that large groups behave as moving obstacles [38]. [39] found that social groups walk slower, further and maintained closer proximity than non-group members. The presence of social groups influenced other pedestrians to walk faster and at a greater distance (even in counter-flow) in order to avoid moving inside the group. The characteristics of movement and walking configurations of social groups were also found to impact the evacuation processes. [40] reports on a positive effect on the evacuation time when groups are present, as self-ordering processes were observed in the crowd at the exits. However, [41] performed egress experiments in which the presence of groups resulted in longer egress times, as members of social groups took longer to respond and move in the direction of the exits. The findings of the studies presented above highlight the influence of social groups on the dynamic of crowds. It is expected that this also applies in the context of railway stations where both moving and waiting passengers are present. It is therefore of great interest to examine the influence of social groups on the capacity of pedestrian facilities such as railway platforms as well as in bottleneck situations like in the boarding and alighting process. Moreover, pedestrians that are members of social groups are expected to use the available space differently. A current application of the results of such an analysis is the detection of offenders against the social distancing rules during the COVID-19 pandemic. While members of social groups, e.g., families, are allowed to have close contact, strangers are obliged to keep a distance from one another in order to reduce the risk of infection. To identify situations or regions in which the mandatory distance is not kept, the identification of social groups and individuals is essential. This paper seeks to fill the gap between existing studies on pedestrian waiting behaviour and research on social groups. Since only limited research has been published concerning the detection and analysis of characteristics of non-walking social groups, this article presents a method to identify social groups at train station platforms, where both moving and waiting / standing behaviour is present. Based on the proposed method of group detection, the waiting behaviour of pedestrians at train station platforms is analysed. The use of trajectory data, in contrast to video recordings, ensures a privacy conserving methodology. The method is applied to data from a railway platform in Switzerland. The portions of pedestrians travelling in groups and the distribution of group sizes are analysed and the differences between members of social groups and individuals are discussed with respect to mean speed and choice of waiting places.

Data sources The tracking data used in this study was provided by the Swiss Federal Railways (SBB AG) and was collected at platform 2/3 of the station Zürich Hardbrücke, Switzerland. This train station platform is equipped with stereo sensors tracking the movement of pedestrians inside the area of observation with 10 frames per second. The data consists of an unique ID number for each pedestrian, a timestamp and the x and y coordinates of the pedestrian’s position at the given timestamp. As only the trajectories are recorded the data is fully privacy conserving and no information can be accessed that would allow to identify any individual pedestrian. The data was collected between 1st and 28th of February 2020 during the afternoon peak hours from 4 p.m. to 7 p.m. The data set thus consists of 8 weekend days and 20 workdays. The chosen time interval does not intersect with any measures introduced during the Covid-19 pandemic. The afternoon peak hours were selected with respect to comparability due to the fact that in these hours the passenger amount is usually high during both workdays and weekends and the most passengers travelling in social groups were expected. While the morning peak hours are often assigned to individual travel to e.g. work places, in the afternoon peak hours social activities are more likely. The observed area covers about 50 metres, see Fig 1. The platform is constructed symmetrically with an information board in the central area and stairways and elevators to both sides. The direction of movement of passengers entering the platform is indicated with arrows at the stairs and elevators in Fig 1. PPT PowerPoint slide

PNG larger image

TIFF original image Download: Fig 1. Spatial structure of the measurement area at the railway platform of Zürich Hardbrücke. Arrows indicate the movement direction for pedestrians entering the platform. The measurement area covers approximately 50 m of the platform. https://doi.org/10.1371/journal.pone.0282526.g001 Data quality In order to assess the quality of the data, the starting and ending points of trajectories were checked for plausibility. Due to the setup of the measurement area, trajectories are expected to begin and end either at the platform entrances, the platform-train interfaces or at the sides of the measurement areas. Approximately 90% of all starting points of trajectories fulfil these requirements; the same applies to the ending points. However, in order to have a “complete” trajectory, both beginning and ending are required to be at the expected positions. This is the case for about 75-80% of the trajectories. Incomplete trajectories were caused by pedestrians being lost by the tracking system for one or more frames, as a new ID number is assigned upon re-detection (cf. [42]). Nevertheless, those data are included in the analysis but the probability that the corresponding pedestrians are assigned as members of social groups is decreased. Due to technical reasons, the tracking data is mirrored horizontally.

Methodology In manual field observations or by watching video recordings social groups can be identified by their social interaction, observable through e.g. verbal communication, eye contact and gestures (cf. [26]). This is not possible when working with trajectory data only. Nonetheless there are many research questions where a differentiation between social groups and individuals is crucial. A method on group detection in railway environments was previously introduced by [36] and used to perform contact tracing and analyse the distancing rules during the first phase of the Covid-19 pandemic. Their method uses a sparse graph in which the trajectories of the pedestrians as well as all events in which two pedestrians had a distance smaller than a predefined threshold of 2.5 m are memorized. Hence, the distances between each pair of pedestrians that is present simultaneously at the platform have to be calculated. The calculation of distances between N pedestrians scales with N2 and is therefore very time-consuming for large N [43]. While the distance calculation between all pedestrians is necessary for the analysis of Covid-19 distancing rules, it is not in the context of social group assessment. As members of social groups are expected to keep closer contact to each other than to non-group members, it is not necessary to calculate the distances between all pedestrians that are present at a given time. It is sufficient to determine the persons standing nearest to one another in every frame by applying Delaunay Triangulation which has a complexity (cf. [44, 45]) and is therewith far more time efficient. Therefore, this paper proposes an adapted method to recognise social groups in trajectory data by analysing the distances and the stability of the distances between neighbouring persons which avoids the calculation of N2 distances. Train station platforms are places where boarding and alighting passengers are present. Alighting passengers usually leave the train platforms in a straight path, which makes it almost impossible to determine socially related pedestrians, even if video recordings were available. However, boarding passengers wait for a certain amount of time and can therefore be observed over longer time intervals. Hence, the identification of social groups is restricted to boarding passengers. The categorisation was performed by determining the start and end points of the trajectories: a boarding passengers trajectory starts at a platform entry and ends at a train door. A detailed discussion and analysis of that matter can be found in the authors’ previous work, see [46]. All trajectories that are shorter than 20 seconds were not included in the analysis, as this time interval was identified as minimal observation time needed for visual analysis (see section Parameter Study and Validation). Since the group detection method identifies group members based on their distances from one another, the applicability is limited to low density environments. In consideration of the goal of determining members of social groups, which will be characterised by reasonably small distances, two different thresholds are defined for the distance between two pedestrians. A value of 1.5 metres was chosen as the maximum distance between persons for a contact (d contact ≤ 1.5m) which was also established as social distancing threshold in numerous European countries during the Covid-19 pandemic and is the maximum of the probability density of the pairwise distances in the data set. In order to regard the personal distance a value of 1 m is used (d personal ≤ 1m) as pedestrians that are comfortable to be inside each others personal space over a longer period are most likely related in a social way. Hence, t contact is determined as the number of frames t for which holds (1) and t personal as the number of frames for which is (2) with being the position of pedestrian i at time t, for pedestrian j respectively. In words, t contact translates to the number of frames in which the given pedestrians i and j are at a distance of 1.5 m or less from one another, and t personal as the number frames for which the distance is smaller than 1 metre. Since the two pedestrians i and j do not necessarily have to arrive and depart at the same time, the time in which both pedestrians i and j are inside the measurement area simultaneously, is calculated as (3) with t i,0 representing the first frame and t i,N the last frame in which pedestrian i is inside the measurement area; for pedestrian j respectively. Following [34, 36] the pedestrian pairs identified based on the small distances between them, will be checked for the following relations: (4) (5) The values for α and β are determined in a parameter study in the following section. If both Eqs (4) and (5) are fulfilled, the corresponding pedestrians i and j are considered to belong to the same social group. Groups with more than two members are detected by combination of pairs. Parameter study and validation To determine a suitable parameter set for α and β and to validate the social groups found by the proposed method a ground truth of IDs that are members of social groups was established. To do so, the trajectory data of one example time interval of three hours was visualised as a video with JPSvis, which is the visualisation tool of the software JuPedSim [47]. Two persons, who had no knowledge of the group detection, were asked to individually note all ID numbers of pedestrians, who they believe to be members of social groups. No specific instructions to the determination of groups were given, but both persons were asked for their strategy afterwards. The test persons identified group members based on simultaneous movements, similar waiting locations and close proximity over longer periods of time. It was monitored whether a certain person entered or left the area of observation along with others, or if the person stayed close to others during the time at the platform. A collective change of waiting positions was also used as indication of group affiliation. However, the visual recognition of groups based solely on trajectories is not a trivial procedure. In order to guarantee reliability, the results of the two test persons were compared. The first test person noted 154 IDs as members of social groups, the second 153 IDs. In total 146 IDs were listed by both testers, which means they agreed in 90.7% of the cases. The IDs identified by both persons were used as ground truth for the parameter study. Therefore, all ID numbers that are not part of the 146 IDs found by both testers were considered to be individuals and in case one of those was found by the group detection method, it was assumed to be a false positive. A parameter study was then performed to determine suitable values for α and β. Hence, suitable parameters are determined based on two constrains. The aim was to find a set of values for which a large number of members of social groups can be detected, however, the number of false positive detections should be zero, as those would correspond to pedestrians that are likely individuals but erroneously marked as group members. The values of α and β were varied between 0 and 1 in steps of 0.05. For each set of values, the group detection was performed based on Eqs (4) and (5). All IDs that were identified for a certain parameter set were than checked against the ground truth in order to determine any false positive. The numbers of false positives increased with decreasing values for α and β (cf. darker colours in Fig 2a)). As the aim was to avoid false positives but to find a large number of group members, the numbers of identified group members for the corresponding parameters are illustrated in Fig 2b, with the red area marking the sets of parameters for which the number of false positives is zero. Therefore, the best parameter set will be identified as {α, β} = argmax(N(members), whereN(falsepositive) = 0), which corresponds to the set that produces the maximum number of found groups members within the red area. From this it can be seen that the best results are achieved with α = 0.85 and β = 0.4. In words, pedestrians are assumed to belong to a social group, if they have a distance smaller than 1.5 m to at least one member of the group for 85% of the time that they are simultaneously inside the measurement area and a distance smaller than 1 m for 40% of that time. The work of [36] proposed values of 90% and 40%, which can therewith be confirmed within 5% by the performed parameter study. This parameter combination allows for a maximum of 107 IDs to be correctly identified as group members without any false positives. This corresponds to about 73% of the group members determined by the testers. Considering the overall data quality, which exhibits about 75-80% of complete trajectories (cf. section Data Quality), these results are satisfactory, as incomplete trajectories will interfere with the group detection method and prevent the correct assignment of group membership. PPT PowerPoint slide

PNG larger image

TIFF original image Download: Fig 2. Parameter study. a) Number of false positive in the detection of group members by different values of α and β. White colours indicate that no false positives were found. b) Number of detected members of social groups. Light colours show high numbers of found members. The red area highlights the sets of parameters where the number of false positives in a) is zero. https://doi.org/10.1371/journal.pone.0282526.g002 The method reaches its limits in crowded situations where higher densities are present over longer time intervals. In those cases, the close distance between pedestrians is not necessarily caused by social interaction but rather by limited available space. Due to the distance thresholds of 1 m and 1.5 m, the method can result in incorrect group assignment if the local density exceeds 0.5 − 1.0 1/m2. If those densities remain over longer time intervals, crowding can be mistaken for social relation. The afternoon peak hours analysed in this study do not exhibit densities that exceed these threshold for longer time intervals. This will likely only occur in highly crowded situations, as e.g. in the context of public events. As the introduced method was developed for low density situations, no prediction can currently be made to what extent members of social groups preserve their close proximity in high density environments. It is expected that at least the distance thresholds and the parameters α and β need to be adjusted in order to correctly determine social groups. It may be necessary to include additional criteria, like for example the simultaneous movement of nearby pedestrians. However, the correlation of movement will not allow to expand the group detection to alighting passengers, since their walking paths inside the area of observation are generally similar and too short to allow assessment of group membership. In environments where pedestrians continuously move the correlation of group members can be expected to be higher than to unrelated pedestrians, even in increasing densities. In the context of train station platforms passengers spend most of their time waiting and therefore do not move, which will increase the correlation between (socially) unrelated neighbours in situations of limited available space (e.g. at a fully crowded platform). Speed calculation and waiting The identification of waiting passengers on train platforms using trajectory data can be achieved by analysing their speed of movement or lack thereof. The speed of a pedestrian at a given time is calculated as the movement of a pedestrian in a time interval Δt. With the location of the pedestrian at time t, the speed can be calculated as: (6) In this study Δt = 50 frames, corresponding to 5 seconds, was used. To determine if a given pedestrian can be considered as waiting, a threshold of v i (t) < 0.4 m/s was applied. This threshold was picked as the local minimum of the velocity distribution of the data set, which shows two peaks: One peak at mean speeds of approximately 0.2 m/s; the other at 1.2 m/s. The first peak mainly relates to boarding, the second to alighting passengers. The velocity distribution can be found in the author’s previous work [46].

Conclusion This paper presented a method that allows the identification of social groups in trajectory data of waiting and standing pedestrians. Social groups were identified by thresholds for inter-personal distances that are present over certain time intervals. In the case of a train station platform, two pedestrians are considered as belonging to a social group, if their distance to one another is smaller than 1.5 m for 85% of the time in which they are simultaneously inside the observation area and smaller than 1 m for 40% of that time. The percentages of the time that are needed for pedestrians to be considered as a social group were determined by a parameter study and were checked against a ground truth for validation. The ground truth was established by a visual analysis performed by two independent test persons. However, these parameters need to be reconsidered and validated for different scenarios, e.g. in shopping malls or public gatherings where different dynamics are occurring. The group detection is suitable for scenarios with low densities; the applicability in dense environments cannot be guaranteed as in those cases small distances between pedestrians are caused by congestion or limited available space. The group detection method was applied to a data set taken from the afternoon peak hours during February 2020 in Zürich Hardbrücke, Switzerland. During working days about 9-10% of the pedestrians waiting at the train station platform were members of social groups; the portion increases to up to 20% during weekends. The most frequently observed group size was pairs, each size is less frequent than the next smaller size. Distributions of group sizes showed no correlation to whether it was a working day or weekend day. With increasing group size, the members mean speed decreased. While individuals often waited at the sides of stairs and elevators, social groups were found to be more likely to choose waiting places that provide enough space for members to position themselves in such a way that enables communication within the group. Typically, this is the case in the vicinity of the platform entrances. This behaviour was shown to be more prominent with increasing group size. Moreover, waiting places were influenced by the total waiting time of the passengers. Pedestrians with short waiting times (less than 2 minutes) waited close the entrances. For longer waiting times places at the undisturbed rearward sides of the stairs were used. The waiting places chosen by individuals and groups highlight the different needs in terms of comfort. The waiting places were chosen based on a ranking of the criteria of short walking distances, the direction of the train arrival, undisturbed waiting places and ensured communication. Depending on the types of users and the waiting time those criteria were prioritised differently. Passengers with long waiting times prefer undisturbed waiting places even if the distance was longer. While individuals chose undisturbed waiting places in areas where they do not hinder the movement of others, social groups prioritised the possibility to communicate even if the position was close to the highly frequented entry way. The results could be used to assess the comfort of different types of users by level of service concept including waiting passengers and to optimise space usage at railway platforms by increasing the robustness of performance during peak load by optimising the pedestrian distribution.

Acknowledgments The authors would like to thank the Swiss Federal Railways for providing the data and the student assistants for their help with the visual group assessment.

[END]
---
[1] Url: https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0282526

Published and (C) by PLOS One
Content appears here under this condition or license: Creative Commons - Attribution BY 4.0.

via Magical.Fish Gopher News Feeds:
gopher://magical.fish/1/feeds/news/plosone/

You are viewing proxied material from magical.fish. The copyright of proxied material belongs to its original authors. Any comments or complaints in relation to proxied material should be directed to the original authors of the content concerned. Please see the disclaimer for more details.